Adding HF files

.gitignore

@@ -0,0 +1,145 @@
+# Specific
+/logs*/
+/results/
+/sandbox/
+*.lock
+*.pt
+*.npy
+/example_data/downloaded*
+*.tar
+*.tar.gz
+/discord_sandbox/
+/demo_out/
+token_channel.csv
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+logs/
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+/checkpoints/
+/data/

README.md

@@ -1,8 +1,8 @@
 ---
-title: HMR
-emoji: 📈
-colorFrom: indigo
-colorTo: pink
+title: HMR2.0
+emoji: 🔥
+colorFrom: yellow
+colorTo: yellow
 sdk: gradio
 sdk_version: 3.32.0
 app_file: app.py

README_github.md

@@ -0,0 +1,74 @@
+# 4DHumans: Reconstructing and Tracking Humans with Transformers
+Code repository for the paper:
+**Humans in 4D: Reconstructing and Tracking Humans with Transformers**
+[Shubham Goel](https://people.eecs.berkeley.edu/~shubham-goel/), [Georgios Pavlakos](https://geopavlakos.github.io/), [Jathushan Rajasegaran](http://people.eecs.berkeley.edu/~jathushan/), [Angjoo Kanazawa](https://people.eecs.berkeley.edu/~kanazawa/)<sup>\*</sup>, [Jitendra Malik](http://people.eecs.berkeley.edu/~malik/)<sup>\*</sup>
+arXiv preprint 2023
+[[paper]()] [[project page](https://shubham-goel.github.io/4dhumans/)] [[hugging faces space]()]
+
+![teaser](assets/teaser.png)
+
+## Download dependencies
+Our demo code depends on [detectron2](https://github.com/facebookresearch/detectron2) to detect humans.
+To automatically download this dependency, clone this repo using `--recursive`, or run `git submodule update --init` if you've already cloned the repository. You should see the detectron2 source code at `vendor/detectron2`.
+```bash
+git clone https://github.com/shubham-goel/4D-Humans.git --recursive
+# OR
+git clone https://github.com/shubham-goel/4D-Humans.git
+cd 4D-Humans
+git submodule update --init
+```
+
+## Installation
+We recommend creating a clean [conda](https://docs.conda.io/) environment and installing all dependencies, as follows:
+```bash
+conda env create -f environment.yml
+```
+
+After the installation is complete you can activate the conda environment by running:
+```
+conda activate 4D-humans
+```
+
+## Download checkpoints and SMPL models
+To download the checkpoints and SMPL models, run
+```bash
+./fetch_data.sh
+```
+
+## Run demo on images
+You may now run our demo to 3D reconstruct humans in images using the following command, which will run ViTDet and HMR2.0 on all images in the specified `--img_folder` and save renderings of the reconstructions in `--out_folder`. You can also use the `--side_view` flag to additionally render the side view of the reconstructed mesh. `--batch_size` batches the images together for faster processing.
+```bash
+python demo.py \
+    --img_folder example_data/images \
+    --out_folder demo_out \
+    --batch_size=48 --side_view
+```
+
+## Run demo on videos
+Coming soon.
+
+## Training and evaluation
+Cmoing soon.
+
+## Acknowledgements
+Parts of the code are taken or adapted from the following repos:
+- [ProHMR](https://github.com/nkolot/ProHMR)
+- [SPIN](https://github.com/nkolot/SPIN)
+- [SMPLify-X](https://github.com/vchoutas/smplify-x)
+- [HMR](https://github.com/akanazawa/hmr)
+- [ViTPose](https://github.com/ViTAE-Transformer/ViTPose)
+- [Detectron2](https://github.com/facebookresearch/detectron2)
+
+Additionally, we thank [StabilityAI](https://stability.ai/) for a generous compute grant that enabled this work.
+
+## Citing
+If you find this code useful for your research, please consider citing the following paper:
+
+```
+@article{4DHUMANS,
+    title={Humans in 4{D}: Reconstructing and Tracking Humans with Transformers},
+    author={Goel, Shubham and Pavlakos, Georgios and Rajasegaran, Jathushan and Kanazawa, Angjoo and Malik, Jitendra},
+    journal={arXiv preprint},
+    year={2023}
+}
+```

app.py

@@ -0,0 +1,149 @@
+import argparse
+import os
+from pathlib import Path
+
+import cv2
+import gradio as gr
+import numpy as np
+import torch
+from PIL import Image
+
+from hmr2.configs import get_config
+from hmr2.datasets.vitdet_dataset import (DEFAULT_MEAN, DEFAULT_STD,
+                                          ViTDetDataset)
+from hmr2.models import HMR2
+from hmr2.utils import recursive_to
+from hmr2.utils.renderer import Renderer, cam_crop_to_full
+
+# Setup HMR2.0 model
+LIGHT_BLUE=(0.65098039,  0.74117647,  0.85882353)
+DEFAULT_CHECKPOINT='logs/train/multiruns/hmr2/0/checkpoints/epoch=35-step=1000000.ckpt'
+device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
+model_cfg = str(Path(DEFAULT_CHECKPOINT).parent.parent / 'model_config.yaml')
+model_cfg = get_config(model_cfg)
+model = HMR2.load_from_checkpoint(DEFAULT_CHECKPOINT, strict=False, cfg=model_cfg).to(device)
+model.eval()
+
+
+# Load detector
+from detectron2.config import LazyConfig
+
+from hmr2.utils.utils_detectron2 import DefaultPredictor_Lazy
+
+detectron2_cfg = LazyConfig.load(f"vendor/detectron2/projects/ViTDet/configs/COCO/cascade_mask_rcnn_vitdet_h_75ep.py")
+detectron2_cfg.train.init_checkpoint = "https://dl.fbaipublicfiles.com/detectron2/ViTDet/COCO/cascade_mask_rcnn_vitdet_h/f328730692/model_final_f05665.pkl"
+for i in range(3):
+    detectron2_cfg.model.roi_heads.box_predictors[i].test_score_thresh = 0.25
+detector = DefaultPredictor_Lazy(detectron2_cfg)
+
+# Setup the renderer
+renderer = Renderer(model_cfg, faces=model.smpl.faces)
+
+
+import numpy as np
+
+
+def infer(in_pil_img, in_threshold=0.8, out_pil_img=None):
+
+    open_cv_image = np.array(in_pil_img)
+    # Convert RGB to BGR
+    open_cv_image = open_cv_image[:, :, ::-1].copy()
+    print("EEEEE", open_cv_image.shape)
+    det_out = detector(open_cv_image)
+    det_instances = det_out['instances']
+    valid_idx = (det_instances.pred_classes==0) & (det_instances.scores > in_threshold)
+    boxes=det_instances.pred_boxes.tensor[valid_idx].cpu().numpy()
+
+    # Run HMR2.0 on all detected humans
+    dataset = ViTDetDataset(model_cfg, open_cv_image, boxes)
+    dataloader = torch.utils.data.DataLoader(dataset, batch_size=8, shuffle=False, num_workers=0)
+
+    all_verts = []
+    all_cam_t = []
+
+    for batch in dataloader:
+        batch = recursive_to(batch, device)
+        with torch.no_grad():
+            out = model(batch)
+
+        pred_cam = out['pred_cam']
+        box_center = batch["box_center"].float()
+        box_size = batch["box_size"].float()
+        img_size = batch["img_size"].float()
+        render_size = img_size
+        pred_cam_t = cam_crop_to_full(pred_cam, box_center, box_size, render_size).detach().cpu().numpy()
+
+        # Render the result
+        batch_size = batch['img'].shape[0]
+        for n in range(batch_size):
+            # Get filename from path img_path
+            # img_fn, _ = os.path.splitext(os.path.basename(img_path))
+            person_id = int(batch['personid'][n])
+            white_img = (torch.ones_like(batch['img'][n]).cpu() - DEFAULT_MEAN[:,None,None]/255) / (DEFAULT_STD[:,None,None]/255)
+            input_patch = batch['img'][n].cpu() * (DEFAULT_STD[:,None,None]/255) + (DEFAULT_MEAN[:,None,None]/255)
+            input_patch = input_patch.permute(1,2,0).numpy()
+
+            regression_img = renderer(out['pred_vertices'][n].detach().cpu().numpy(),
+                                    out['pred_cam_t'][n].detach().cpu().numpy(),
+                                    batch['img'][n],
+                                    mesh_base_color=LIGHT_BLUE,
+                                    scene_bg_color=(1, 1, 1),
+                                    )
+
+
+            verts = out['pred_vertices'][n].detach().cpu().numpy()
+            cam_t = pred_cam_t[n]
+
+            all_verts.append(verts)
+            all_cam_t.append(cam_t)
+
+
+    # Render front view
+    if len(all_verts) > 0:
+        misc_args = dict(
+            mesh_base_color=LIGHT_BLUE,
+            scene_bg_color=(1, 1, 1),
+        )
+        cam_view = renderer.render_rgba_multiple(all_verts, cam_t=all_cam_t, render_res=render_size[n], **misc_args)
+
+        # Overlay image
+        input_img = open_cv_image.astype(np.float32)[:,:,::-1]/255.0
+        input_img = np.concatenate([input_img, np.ones_like(input_img[:,:,:1])], axis=2) # Add alpha channel
+        input_img_overlay = input_img[:,:,:3] * (1-cam_view[:,:,3:]) + cam_view[:,:,:3] * cam_view[:,:,3:]
+
+        # convert to PIL image
+        out_pil_img =  Image.fromarray((input_img_overlay*255).astype(np.uint8))
+
+        return out_pil_img
+    else:
+        return None
+
+
+with gr.Blocks(title="4DHumans", css=".gradio-container") as demo:
+
+    gr.HTML("""<div style="font-weight:bold; text-align:center; color:royalblue;">HMR 2.0</div>""")
+
+    with gr.Row():
+        input_image = gr.Image(label="Input image", type="pil", width=300, height=300, fixed_size=True)
+        output_image = gr.Image(label="Reconstructions", type="pil", width=300, height=300, fixed_size=True)
+
+    gr.HTML("""<br/>""")
+
+    with gr.Row():
+        threshold = gr.Slider(0, 1.0, value=0.8, label='Detection Threshold')
+        send_btn = gr.Button("Infer")
+        send_btn.click(fn=infer, inputs=[input_image, threshold], outputs=[output_image])
+
+    # gr.Examples(['samples/img1.jpg', 'samples/img2.png', 'samples/img3.jpg', 'samples/img4.jpg'], inputs=input_image)
+
+    gr.HTML("""</ul>""")
+
+
+
+#demo.queue()
+demo.launch(debug=True)
+
+
+
+
+### EOF ###

app_mesh.py

@@ -0,0 +1,141 @@
+import argparse
+import os
+from pathlib import Path
+
+import cv2
+import gradio as gr
+import numpy as np
+import torch
+from PIL import Image
+import trimesh
+import tempfile
+
+from hmr2.configs import get_config
+from hmr2.datasets.vitdet_dataset import (DEFAULT_MEAN, DEFAULT_STD,
+                                          ViTDetDataset)
+from hmr2.models import HMR2
+from hmr2.utils import recursive_to
+from hmr2.utils.renderer import Renderer, cam_crop_to_full
+
+# Setup HMR2.0 model
+LIGHT_BLUE=(0.65098039,  0.74117647,  0.85882353)
+DEFAULT_CHECKPOINT='logs/train/multiruns/hmr2/0/checkpoints/epoch=35-step=1000000.ckpt'
+device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
+model_cfg = str(Path(DEFAULT_CHECKPOINT).parent.parent / 'model_config.yaml')
+model_cfg = get_config(model_cfg)
+model = HMR2.load_from_checkpoint(DEFAULT_CHECKPOINT, strict=False, cfg=model_cfg).to(device)
+model.eval()
+
+
+# Load detector
+from detectron2.config import LazyConfig
+
+from hmr2.utils.utils_detectron2 import DefaultPredictor_Lazy
+
+detectron2_cfg = LazyConfig.load(f"vendor/detectron2/projects/ViTDet/configs/COCO/cascade_mask_rcnn_vitdet_h_75ep.py")
+detectron2_cfg.train.init_checkpoint = "https://dl.fbaipublicfiles.com/detectron2/ViTDet/COCO/cascade_mask_rcnn_vitdet_h/f328730692/model_final_f05665.pkl"
+for i in range(3):
+    detectron2_cfg.model.roi_heads.box_predictors[i].test_score_thresh = 0.25
+detector = DefaultPredictor_Lazy(detectron2_cfg)
+
+# Setup the renderer
+renderer = Renderer(model_cfg, faces=model.smpl.faces)
+
+
+import numpy as np
+
+
+def infer(in_pil_img, in_threshold=0.8):
+
+    open_cv_image = np.array(in_pil_img)
+    # Convert RGB to BGR
+    open_cv_image = open_cv_image[:, :, ::-1].copy()
+    print("EEEEE", open_cv_image.shape)
+    det_out = detector(open_cv_image)
+    det_instances = det_out['instances']
+    valid_idx = (det_instances.pred_classes==0) & (det_instances.scores > in_threshold)
+    boxes=det_instances.pred_boxes.tensor[valid_idx].cpu().numpy()
+
+    # Run HMR2.0 on all detected humans
+    dataset = ViTDetDataset(model_cfg, open_cv_image, boxes)
+    dataloader = torch.utils.data.DataLoader(dataset, batch_size=8, shuffle=False, num_workers=0)
+
+    all_verts = []
+    all_cam_t = []
+
+    for batch in dataloader:
+        batch = recursive_to(batch, device)
+        with torch.no_grad():
+            out = model(batch)
+
+        pred_cam = out['pred_cam']
+        box_center = batch["box_center"].float()
+        box_size = batch["box_size"].float()
+        img_size = batch["img_size"].float()
+        render_size = img_size
+        pred_cam_t = cam_crop_to_full(pred_cam, box_center, box_size, render_size, focal_length=img_size.mean()*2).detach().cpu().numpy()
+
+        # Render the result
+        batch_size = batch['img'].shape[0]
+        for n in range(batch_size):
+            # Get filename from path img_path
+            # img_fn, _ = os.path.splitext(os.path.basename(img_path))
+            person_id = int(batch['personid'][n])
+            white_img = (torch.ones_like(batch['img'][n]).cpu() - DEFAULT_MEAN[:,None,None]/255) / (DEFAULT_STD[:,None,None]/255)
+            input_patch = batch['img'][n].cpu() * (DEFAULT_STD[:,None,None]/255) + (DEFAULT_MEAN[:,None,None]/255)
+            input_patch = input_patch.permute(1,2,0).numpy()
+
+            regression_img = renderer(out['pred_vertices'][n].detach().cpu().numpy(),
+                                    out['pred_cam_t'][n].detach().cpu().numpy(),
+                                    batch['img'][n],
+                                    mesh_base_color=LIGHT_BLUE,
+                                    scene_bg_color=(1, 1, 1),
+                                    )
+
+
+            verts = out['pred_vertices'][n].detach().cpu().numpy()
+            cam_t = pred_cam_t[n]
+
+            all_verts.append(verts)
+            all_cam_t.append(cam_t)
+
+    # Return mesh path
+    trimeshes = [renderer.vertices_to_trimesh(vvv, ttt.copy(), LIGHT_BLUE) for vvv,ttt in zip(all_verts, all_cam_t)]
+
+    # Join meshes
+    mesh = trimesh.util.concatenate(trimeshes)
+
+    # Save mesh to file
+    temp_name = next(tempfile._get_candidate_names()) + '.obj'
+    trimesh.exchange.export.export_mesh(mesh, temp_name)
+    return temp_name
+
+
+with gr.Blocks(title="4DHumans", css=".gradio-container") as demo:
+
+    gr.HTML("""<div style="font-weight:bold; text-align:center; color:royalblue;">HMR 2.0</div>""")
+
+    with gr.Row():
+        input_image = gr.Image(label="Input image", type="pil", width=300, height=300, fixed_size=True)
+        output_model = gr.Model3D(label="Reconstructions", width=300, height=300, fixed_size=True, clear_color=[0.0, 0.0, 0.0, 0.0])
+
+    gr.HTML("""<br/>""")
+
+    with gr.Row():
+        threshold = gr.Slider(0, 1.0, value=0.8, label='Detection Threshold')
+        send_btn = gr.Button("Infer")
+        send_btn.click(fn=infer, inputs=[input_image, threshold], outputs=[output_model])
+
+    # gr.Examples(['samples/img1.jpg', 'samples/img2.png', 'samples/img3.jpg', 'samples/img4.jpg'], inputs=input_image)
+
+    gr.HTML("""</ul>""")
+
+
+
+#demo.queue()
+demo.launch(debug=True)
+
+
+
+
+### EOF ###

clean_ckpt.ipynb

@@ -0,0 +1,93 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "ckpt_path = 'logs/train/multiruns/hmr2/0/checkpoints/epoch=35-step=1000000.ckpt'"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Load ckpt\n",
+    "ckpt = torch.load(ckpt_path, map_location='cpu')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "ckpt.keys()\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "ckpt['loops']"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Delete optimizer_states\n",
+    "del ckpt['optimizer_states']\n",
+    "del ckpt['callbacks']\n",
+    "del ckpt['loops']"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Save new ckpt\n",
+    "torch.save(ckpt, ckpt_path)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "4D-humans",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.6"
+  },
+  "orig_nbformat": 4
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

demo.py

@@ -0,0 +1,130 @@
+from pathlib import Path
+import torch
+import argparse
+import os
+import cv2
+import numpy as np
+
+from hmr2.configs import get_config
+from hmr2.models import HMR2
+from hmr2.utils import recursive_to
+from hmr2.datasets.vitdet_dataset import ViTDetDataset, DEFAULT_MEAN, DEFAULT_STD
+from hmr2.utils.renderer import Renderer, cam_crop_to_full
+
+LIGHT_BLUE=(0.65098039,  0.74117647,  0.85882353)
+# DEFAULT_CHECKPOINT='logs/train/multiruns/20b1_mix11_a1/0/checkpoints/epoch=30-step=1000000.ckpt'
+DEFAULT_CHECKPOINT='logs/train/multiruns/hmr2/0/checkpoints/epoch=35-step=1000000.ckpt'
+parser = argparse.ArgumentParser(description='HMR2 demo code')
+parser.add_argument('--checkpoint', type=str, default=DEFAULT_CHECKPOINT, help='Path to pretrained model checkpoint')
+parser.add_argument('--img_folder', type=str, default='example_data/images', help='Folder with input images')
+parser.add_argument('--out_folder', type=str, default='demo_out', help='Output folder to save rendered results')
+parser.add_argument('--side_view', dest='side_view', action='store_true', default=False, help='If set, render side view also')
+parser.add_argument('--batch_size', type=int, default=1, help='Batch size for inference/fitting')
+
+args = parser.parse_args()
+
+# Setup HMR2.0 model
+device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
+model_cfg = str(Path(args.checkpoint).parent.parent / 'model_config.yaml')
+model_cfg = get_config(model_cfg)
+model = HMR2.load_from_checkpoint(args.checkpoint, strict=False, cfg=model_cfg).to(device)
+model.eval()
+
+# Load detector
+from detectron2.config import LazyConfig
+from hmr2.utils.utils_detectron2 import DefaultPredictor_Lazy
+detectron2_cfg = LazyConfig.load(f"vendor/detectron2/projects/ViTDet/configs/COCO/cascade_mask_rcnn_vitdet_h_75ep.py")
+detectron2_cfg.train.init_checkpoint = "https://dl.fbaipublicfiles.com/detectron2/ViTDet/COCO/cascade_mask_rcnn_vitdet_h/f328730692/model_final_f05665.pkl"
+for i in range(3):
+    detectron2_cfg.model.roi_heads.box_predictors[i].test_score_thresh = 0.25
+detector = DefaultPredictor_Lazy(detectron2_cfg)
+
+# Setup the renderer
+renderer = Renderer(model_cfg, faces=model.smpl.faces)
+
+# Make output directory if it does not exist
+os.makedirs(args.out_folder, exist_ok=True)
+
+# Iterate over all images in folder
+for img_path in Path(args.img_folder).glob('*.png'):
+    img_cv2 = cv2.imread(str(img_path), cv2.IMREAD_COLOR)
+
+    # Detect humans in image
+    det_out = detector(img_cv2)
+
+    det_instances = det_out['instances']
+    valid_idx = (det_instances.pred_classes==0) & (det_instances.scores > 0.5)
+    boxes=det_instances.pred_boxes.tensor[valid_idx].cpu().numpy()
+
+    # Run HMR2.0 on all detected humans
+    dataset = ViTDetDataset(model_cfg, img_cv2.copy(), boxes)
+    dataloader = torch.utils.data.DataLoader(dataset, batch_size=8, shuffle=False, num_workers=0)
+
+
+    all_verts = []
+    all_cam_t = []
+
+    for batch in dataloader:
+        batch = recursive_to(batch, device)
+        with torch.no_grad():
+            out = model(batch)
+
+        pred_cam = out['pred_cam']
+        box_center = batch["box_center"].float()
+        box_size = batch["box_size"].float()
+        img_size = batch["img_size"].float()
+        render_size = img_size
+        pred_cam_t = cam_crop_to_full(pred_cam, box_center, box_size, render_size).detach().cpu().numpy()
+
+        # Render the result
+        batch_size = batch['img'].shape[0]
+        for n in range(batch_size):
+            # Get filename from path img_path
+            img_fn, _ = os.path.splitext(os.path.basename(img_path))
+            person_id = int(batch['personid'][n])
+            white_img = (torch.ones_like(batch['img'][n]).cpu() - DEFAULT_MEAN[:,None,None]/255) / (DEFAULT_STD[:,None,None]/255)
+            input_patch = batch['img'][n].cpu() * (DEFAULT_STD[:,None,None]/255) + (DEFAULT_MEAN[:,None,None]/255)
+            input_patch = input_patch.permute(1,2,0).numpy()
+            
+            regression_img = renderer(out['pred_vertices'][n].detach().cpu().numpy(),
+                                    out['pred_cam_t'][n].detach().cpu().numpy(),
+                                    batch['img'][n],
+                                    mesh_base_color=LIGHT_BLUE,
+                                    scene_bg_color=(1, 1, 1),
+                                    )
+
+            if args.side_view:
+                side_img = renderer(out['pred_vertices'][n].detach().cpu().numpy(),
+                                        out['pred_cam_t'][n].detach().cpu().numpy(),
+                                        white_img,
+                                        mesh_base_color=LIGHT_BLUE,
+                                        scene_bg_color=(1, 1, 1),
+                                        side_view=True)
+                final_img = np.concatenate([input_patch, regression_img, side_img], axis=1)
+            else:
+                final_img = np.concatenate([input_patch, regression_img], axis=1)
+
+
+            verts = out['pred_vertices'][n].detach().cpu().numpy()
+            cam_t = pred_cam_t[n]
+
+            all_verts.append(verts)
+            all_cam_t.append(cam_t)
+
+    misc_args = dict(
+        mesh_base_color=LIGHT_BLUE,
+        scene_bg_color=(1, 1, 1),
+    )
+
+    # Render front view
+    if len(all_verts) > 0:
+        cam_view = renderer.render_rgba_multiple(all_verts, cam_t=all_cam_t, render_res=render_size[n], **misc_args)
+
+        # Overlay image
+        input_img = img_cv2.astype(np.float32)[:,:,::-1]/255.0
+        input_img = np.concatenate([input_img, np.ones_like(input_img[:,:,:1])], axis=2) # Add alpha channel
+        input_img_overlay = input_img[:,:,:3] * (1-cam_view[:,:,3:]) + cam_view[:,:,:3] * cam_view[:,:,3:]
+
+
+        # cv2.imwrite(os.path.join(args.out_folder, f'{img_fn}_{person_id}.jpg'), 255*final_img[:, :, ::-1])
+        cv2.imwrite(os.path.join(args.out_folder, f'rend_{img_fn}.jpg'), 255*input_img_overlay[:, :, ::-1])

environment.yml

@@ -0,0 +1,20 @@
+name: 4D-humans
+channels:
+  - pytorch
+  - nvidia
+  - conda-forge
+dependencies:
+  - python=3.10
+  - pytorch-cuda=11.8
+  - torchvision
+  - pip
+  - pip:
+    - pytorch-lightning
+    - smplx==0.1.28
+    - pyrender
+    - opencv-python
+    - yacs
+    - scikit-image
+    - einops
+    - timm
+    - -e ./vendor/detectron2/

fetch_data.sh

@@ -0,0 +1,2 @@
+wget https://people.eecs.berkeley.edu/~shubham-goel/projects/4DHumans/hmr2_data.tar.gz
+tar -xvzf hmr2_data.tar.gz

hmr2/configs/__init__.py

@@ -0,0 +1,88 @@
+import os
+from typing import Dict
+from yacs.config import CfgNode as CN
+
+
+def to_lower(x: Dict) -> Dict:
+    """
+    Convert all dictionary keys to lowercase
+    Args:
+      x (dict): Input dictionary
+    Returns:
+      dict: Output dictionary with all keys converted to lowercase
+    """
+    return {k.lower(): v for k, v in x.items()}
+
+_C = CN(new_allowed=True)
+
+_C.GENERAL = CN(new_allowed=True)
+_C.GENERAL.RESUME = True
+_C.GENERAL.TIME_TO_RUN = 3300
+_C.GENERAL.VAL_STEPS = 100
+_C.GENERAL.LOG_STEPS = 100
+_C.GENERAL.CHECKPOINT_STEPS = 20000
+_C.GENERAL.CHECKPOINT_DIR = "checkpoints"
+_C.GENERAL.SUMMARY_DIR = "tensorboard"
+_C.GENERAL.NUM_GPUS = 1
+_C.GENERAL.NUM_WORKERS = 4
+_C.GENERAL.MIXED_PRECISION = True
+_C.GENERAL.ALLOW_CUDA = True
+_C.GENERAL.PIN_MEMORY = False
+_C.GENERAL.DISTRIBUTED = False
+_C.GENERAL.LOCAL_RANK = 0
+_C.GENERAL.USE_SYNCBN = False
+_C.GENERAL.WORLD_SIZE = 1
+
+_C.TRAIN = CN(new_allowed=True)
+_C.TRAIN.NUM_EPOCHS = 100
+_C.TRAIN.BATCH_SIZE = 32
+_C.TRAIN.SHUFFLE = True
+_C.TRAIN.WARMUP = False
+_C.TRAIN.NORMALIZE_PER_IMAGE = False
+_C.TRAIN.CLIP_GRAD = False
+_C.TRAIN.CLIP_GRAD_VALUE = 1.0
+_C.LOSS_WEIGHTS = CN(new_allowed=True)
+
+_C.DATASETS = CN(new_allowed=True)
+
+_C.MODEL = CN(new_allowed=True)
+_C.MODEL.IMAGE_SIZE = 224
+
+_C.EXTRA = CN(new_allowed=True)
+_C.EXTRA.FOCAL_LENGTH = 5000
+
+_C.DATASETS.CONFIG = CN(new_allowed=True)
+_C.DATASETS.CONFIG.SCALE_FACTOR = 0.3
+_C.DATASETS.CONFIG.ROT_FACTOR = 30
+_C.DATASETS.CONFIG.TRANS_FACTOR = 0.02
+_C.DATASETS.CONFIG.COLOR_SCALE = 0.2
+_C.DATASETS.CONFIG.ROT_AUG_RATE = 0.6
+_C.DATASETS.CONFIG.TRANS_AUG_RATE = 0.5
+_C.DATASETS.CONFIG.DO_FLIP = True
+_C.DATASETS.CONFIG.FLIP_AUG_RATE = 0.5
+_C.DATASETS.CONFIG.EXTREME_CROP_AUG_RATE = 0.10
+
+def default_config() -> CN:
+    """
+    Get a yacs CfgNode object with the default config values.
+    """
+    # Return a clone so that the defaults will not be altered
+    # This is for the "local variable" use pattern
+    return _C.clone()
+
+def get_config(config_file: str, merge: bool = True) -> CN:
+    """
+    Read a config file and optionally merge it with the default config file.
+    Args:
+      config_file (str): Path to config file.
+      merge (bool): Whether to merge with the default config or not.
+    Returns:
+      CfgNode: Config as a yacs CfgNode object.
+    """
+    if merge:
+      cfg = default_config()
+    else:
+      cfg = CN(new_allowed=True)
+    cfg.merge_from_file(config_file)
+    cfg.freeze()
+    return cfg

hmr2/datasets/utils.py

@@ -0,0 +1,999 @@
+"""
+Parts of the code are taken or adapted from
+https://github.com/mkocabas/EpipolarPose/blob/master/lib/utils/img_utils.py
+"""
+import torch
+import numpy as np
+from skimage.transform import rotate, resize
+from skimage.filters import gaussian
+import random
+import cv2
+from typing import List, Dict, Tuple
+from yacs.config import CfgNode
+
+def expand_to_aspect_ratio(input_shape, target_aspect_ratio=None):
+    """Increase the size of the bounding box to match the target shape."""
+    if target_aspect_ratio is None:
+        return input_shape
+
+    try:
+        w , h = input_shape
+    except (ValueError, TypeError):
+        return input_shape
+
+    w_t, h_t = target_aspect_ratio
+    if h / w < h_t / w_t:
+        h_new = max(w * h_t / w_t, h)
+        w_new = w
+    else:
+        h_new = h
+        w_new = max(h * w_t / h_t, w)
+    if h_new < h or w_new < w:
+        breakpoint()
+    return np.array([w_new, h_new])
+
+def do_augmentation(aug_config: CfgNode) -> Tuple:
+    """
+    Compute random augmentation parameters.
+    Args:
+        aug_config (CfgNode): Config containing augmentation parameters.
+    Returns:
+        scale (float): Box rescaling factor.
+        rot (float): Random image rotation.
+        do_flip (bool): Whether to flip image or not.
+        do_extreme_crop (bool): Whether to apply extreme cropping (as proposed in EFT).
+        color_scale (List): Color rescaling factor
+        tx (float): Random translation along the x axis.
+        ty (float): Random translation along the y axis. 
+    """
+
+    tx = np.clip(np.random.randn(), -1.0, 1.0) * aug_config.TRANS_FACTOR
+    ty = np.clip(np.random.randn(), -1.0, 1.0) * aug_config.TRANS_FACTOR
+    scale = np.clip(np.random.randn(), -1.0, 1.0) * aug_config.SCALE_FACTOR + 1.0
+    rot = np.clip(np.random.randn(), -2.0,
+                  2.0) * aug_config.ROT_FACTOR if random.random() <= aug_config.ROT_AUG_RATE else 0
+    do_flip = aug_config.DO_FLIP and random.random() <= aug_config.FLIP_AUG_RATE
+    do_extreme_crop = random.random() <= aug_config.EXTREME_CROP_AUG_RATE
+    extreme_crop_lvl = aug_config.get('EXTREME_CROP_AUG_LEVEL', 0)
+    # extreme_crop_lvl = 0
+    c_up = 1.0 + aug_config.COLOR_SCALE
+    c_low = 1.0 - aug_config.COLOR_SCALE
+    color_scale = [random.uniform(c_low, c_up), random.uniform(c_low, c_up), random.uniform(c_low, c_up)]
+    return scale, rot, do_flip, do_extreme_crop, extreme_crop_lvl, color_scale, tx, ty
+
+def rotate_2d(pt_2d: np.array, rot_rad: float) -> np.array:
+    """
+    Rotate a 2D point on the x-y plane.
+    Args:
+        pt_2d (np.array): Input 2D point with shape (2,).
+        rot_rad (float): Rotation angle
+    Returns:
+        np.array: Rotated 2D point.
+    """
+    x = pt_2d[0]
+    y = pt_2d[1]
+    sn, cs = np.sin(rot_rad), np.cos(rot_rad)
+    xx = x * cs - y * sn
+    yy = x * sn + y * cs
+    return np.array([xx, yy], dtype=np.float32)
+
+
+def gen_trans_from_patch_cv(c_x: float, c_y: float,
+                            src_width: float, src_height: float,
+                            dst_width: float, dst_height: float,
+                            scale: float, rot: float) -> np.array:
+    """
+    Create transformation matrix for the bounding box crop.
+    Args:
+        c_x (float): Bounding box center x coordinate in the original image.
+        c_y (float): Bounding box center y coordinate in the original image.
+        src_width (float): Bounding box width.
+        src_height (float): Bounding box height.
+        dst_width (float): Output box width.
+        dst_height (float): Output box height.
+        scale (float): Rescaling factor for the bounding box (augmentation).
+        rot (float): Random rotation applied to the box.
+    Returns:
+        trans (np.array): Target geometric transformation.
+    """
+    # augment size with scale
+    src_w = src_width * scale
+    src_h = src_height * scale
+    src_center = np.zeros(2)
+    src_center[0] = c_x
+    src_center[1] = c_y
+    # augment rotation
+    rot_rad = np.pi * rot / 180
+    src_downdir = rotate_2d(np.array([0, src_h * 0.5], dtype=np.float32), rot_rad)
+    src_rightdir = rotate_2d(np.array([src_w * 0.5, 0], dtype=np.float32), rot_rad)
+
+    dst_w = dst_width
+    dst_h = dst_height
+    dst_center = np.array([dst_w * 0.5, dst_h * 0.5], dtype=np.float32)
+    dst_downdir = np.array([0, dst_h * 0.5], dtype=np.float32)
+    dst_rightdir = np.array([dst_w * 0.5, 0], dtype=np.float32)
+
+    src = np.zeros((3, 2), dtype=np.float32)
+    src[0, :] = src_center
+    src[1, :] = src_center + src_downdir
+    src[2, :] = src_center + src_rightdir
+
+    dst = np.zeros((3, 2), dtype=np.float32)
+    dst[0, :] = dst_center
+    dst[1, :] = dst_center + dst_downdir
+    dst[2, :] = dst_center + dst_rightdir
+
+    trans = cv2.getAffineTransform(np.float32(src), np.float32(dst))
+
+    return trans
+
+
+def trans_point2d(pt_2d: np.array, trans: np.array):
+    """
+    Transform a 2D point using translation matrix trans.
+    Args:
+        pt_2d (np.array): Input 2D point with shape (2,).
+        trans (np.array): Transformation matrix.
+    Returns:
+        np.array: Transformed 2D point.
+    """
+    src_pt = np.array([pt_2d[0], pt_2d[1], 1.]).T
+    dst_pt = np.dot(trans, src_pt)
+    return dst_pt[0:2]
+
+def get_transform(center, scale, res, rot=0):
+    """Generate transformation matrix."""
+    """Taken from PARE: https://github.com/mkocabas/PARE/blob/6e0caca86c6ab49ff80014b661350958e5b72fd8/pare/utils/image_utils.py"""
+    h = 200 * scale
+    t = np.zeros((3, 3))
+    t[0, 0] = float(res[1]) / h
+    t[1, 1] = float(res[0]) / h
+    t[0, 2] = res[1] * (-float(center[0]) / h + .5)
+    t[1, 2] = res[0] * (-float(center[1]) / h + .5)
+    t[2, 2] = 1
+    if not rot == 0:
+        rot = -rot  # To match direction of rotation from cropping
+        rot_mat = np.zeros((3, 3))
+        rot_rad = rot * np.pi / 180
+        sn, cs = np.sin(rot_rad), np.cos(rot_rad)
+        rot_mat[0, :2] = [cs, -sn]
+        rot_mat[1, :2] = [sn, cs]
+        rot_mat[2, 2] = 1
+        # Need to rotate around center
+        t_mat = np.eye(3)
+        t_mat[0, 2] = -res[1] / 2
+        t_mat[1, 2] = -res[0] / 2
+        t_inv = t_mat.copy()
+        t_inv[:2, 2] *= -1
+        t = np.dot(t_inv, np.dot(rot_mat, np.dot(t_mat, t)))
+    return t
+
+
+def transform(pt, center, scale, res, invert=0, rot=0, as_int=True):
+    """Transform pixel location to different reference."""
+    """Taken from PARE: https://github.com/mkocabas/PARE/blob/6e0caca86c6ab49ff80014b661350958e5b72fd8/pare/utils/image_utils.py"""
+    t = get_transform(center, scale, res, rot=rot)
+    if invert:
+        t = np.linalg.inv(t)
+    new_pt = np.array([pt[0] - 1, pt[1] - 1, 1.]).T
+    new_pt = np.dot(t, new_pt)
+    if as_int:
+        new_pt = new_pt.astype(int)
+    return new_pt[:2] + 1
+
+def crop_img(img, ul, br, border_mode=cv2.BORDER_CONSTANT, border_value=0):
+    c_x = (ul[0] + br[0])/2
+    c_y = (ul[1] + br[1])/2
+    bb_width = patch_width = br[0] - ul[0]
+    bb_height = patch_height = br[1] - ul[1]
+    trans = gen_trans_from_patch_cv(c_x, c_y, bb_width, bb_height, patch_width, patch_height, 1.0, 0)
+    img_patch = cv2.warpAffine(img, trans, (int(patch_width), int(patch_height)), 
+                                flags=cv2.INTER_LINEAR, 
+                                borderMode=border_mode,
+                                borderValue=border_value
+                        )
+    
+    # Force borderValue=cv2.BORDER_CONSTANT for alpha channel
+    if (img.shape[2] == 4) and (border_mode != cv2.BORDER_CONSTANT):
+        img_patch[:,:,3] = cv2.warpAffine(img[:,:,3], trans, (int(patch_width), int(patch_height)), 
+                                            flags=cv2.INTER_LINEAR, 
+                                            borderMode=cv2.BORDER_CONSTANT,
+                            )
+
+    return img_patch
+
+def generate_image_patch_skimage(img: np.array, c_x: float, c_y: float,
+                                 bb_width: float, bb_height: float,
+                                 patch_width: float, patch_height: float,
+                                 do_flip: bool, scale: float, rot: float,
+                                 border_mode=cv2.BORDER_CONSTANT, border_value=0) -> Tuple[np.array, np.array]:
+    """
+    Crop image according to the supplied bounding box.
+    Args:
+        img (np.array): Input image of shape (H, W, 3)
+        c_x (float): Bounding box center x coordinate in the original image.
+        c_y (float): Bounding box center y coordinate in the original image.
+        bb_width (float): Bounding box width.
+        bb_height (float): Bounding box height.
+        patch_width (float): Output box width.
+        patch_height (float): Output box height.
+        do_flip (bool): Whether to flip image or not.
+        scale (float): Rescaling factor for the bounding box (augmentation).
+        rot (float): Random rotation applied to the box.
+    Returns:
+        img_patch (np.array): Cropped image patch of shape (patch_height, patch_height, 3)
+        trans (np.array): Transformation matrix.
+    """
+    
+    img_height, img_width, img_channels = img.shape
+    if do_flip:
+       img = img[:, ::-1, :]
+       c_x = img_width - c_x - 1
+
+    trans = gen_trans_from_patch_cv(c_x, c_y, bb_width, bb_height, patch_width, patch_height, scale, rot)
+
+    #img_patch = cv2.warpAffine(img, trans, (int(patch_width), int(patch_height)), flags=cv2.INTER_LINEAR)
+
+    # skimage
+    center = np.zeros(2)
+    center[0] = c_x
+    center[1] = c_y
+    res = np.zeros(2)
+    res[0] = patch_width
+    res[1] = patch_height
+    # assumes bb_width = bb_height
+    # assumes patch_width = patch_height
+    assert bb_width == bb_height, f'{bb_width=} != {bb_height=}'
+    assert patch_width == patch_height, f'{patch_width=} != {patch_height=}'
+    scale1 = scale*bb_width/200.
+    
+    # Upper left point
+    ul = np.array(transform([1, 1], center, scale1, res, invert=1, as_int=False)) - 1
+    # Bottom right point
+    br = np.array(transform([res[0] + 1,
+                             res[1] + 1], center, scale1, res, invert=1, as_int=False)) - 1
+
+    # Padding so that when rotated proper amount of context is included
+    try:
+        pad = int(np.linalg.norm(br - ul) / 2 - float(br[1] - ul[1]) / 2) + 1
+    except:
+        breakpoint()
+    if not rot == 0:
+        ul -= pad
+        br += pad
+
+
+    if False:
+        # Old way of cropping image
+        ul_int = ul.astype(int)
+        br_int = br.astype(int)
+        new_shape = [br_int[1] - ul_int[1], br_int[0] - ul_int[0]]
+        if len(img.shape) > 2:
+            new_shape += [img.shape[2]]
+        new_img = np.zeros(new_shape)
+
+        # Range to fill new array
+        new_x = max(0, -ul_int[0]), min(br_int[0], len(img[0])) - ul_int[0]
+        new_y = max(0, -ul_int[1]), min(br_int[1], len(img)) - ul_int[1]
+        # Range to sample from original image
+        old_x = max(0, ul_int[0]), min(len(img[0]), br_int[0])
+        old_y = max(0, ul_int[1]), min(len(img), br_int[1])
+        new_img[new_y[0]:new_y[1], new_x[0]:new_x[1]] = img[old_y[0]:old_y[1],
+                                                        old_x[0]:old_x[1]]
+
+    # New way of cropping image
+    new_img = crop_img(img, ul, br, border_mode=border_mode, border_value=border_value).astype(np.float32)
+
+    # print(f'{new_img.shape=}')
+    # print(f'{new_img1.shape=}')
+    # print(f'{np.allclose(new_img, new_img1)=}')
+    # print(f'{img.dtype=}')
+
+
+    if not rot == 0:
+        # Remove padding
+
+        new_img = rotate(new_img, rot) # scipy.misc.imrotate(new_img, rot)
+        new_img = new_img[pad:-pad, pad:-pad]
+
+    if new_img.shape[0] < 1 or new_img.shape[1] < 1:
+        print(f'{img.shape=}')
+        print(f'{new_img.shape=}')
+        print(f'{ul=}')
+        print(f'{br=}')
+        print(f'{pad=}')
+        print(f'{rot=}')
+
+        breakpoint()
+
+    # resize image
+    new_img = resize(new_img, res) # scipy.misc.imresize(new_img, res)
+    
+    new_img = np.clip(new_img, 0, 255).astype(np.uint8)
+
+    return new_img, trans
+
+
+def generate_image_patch_cv2(img: np.array, c_x: float, c_y: float,
+                             bb_width: float, bb_height: float,
+                             patch_width: float, patch_height: float,
+                             do_flip: bool, scale: float, rot: float,
+                             border_mode=cv2.BORDER_CONSTANT, border_value=0) -> Tuple[np.array, np.array]:
+    """
+    Crop the input image and return the crop and the corresponding transformation matrix.
+    Args:
+        img (np.array): Input image of shape (H, W, 3)
+        c_x (float): Bounding box center x coordinate in the original image.
+        c_y (float): Bounding box center y coordinate in the original image.
+        bb_width (float): Bounding box width.
+        bb_height (float): Bounding box height.
+        patch_width (float): Output box width.
+        patch_height (float): Output box height.
+        do_flip (bool): Whether to flip image or not.
+        scale (float): Rescaling factor for the bounding box (augmentation).
+        rot (float): Random rotation applied to the box.
+    Returns:
+        img_patch (np.array): Cropped image patch of shape (patch_height, patch_height, 3)
+        trans (np.array): Transformation matrix.
+    """
+
+    img_height, img_width, img_channels = img.shape
+    if do_flip:
+        img = img[:, ::-1, :]
+        c_x = img_width - c_x - 1
+
+
+    trans = gen_trans_from_patch_cv(c_x, c_y, bb_width, bb_height, patch_width, patch_height, scale, rot)
+
+    img_patch = cv2.warpAffine(img, trans, (int(patch_width), int(patch_height)), 
+                        flags=cv2.INTER_LINEAR, 
+                        borderMode=border_mode,
+                        borderValue=border_value,
+                )
+    # Force borderValue=cv2.BORDER_CONSTANT for alpha channel
+    if (img.shape[2] == 4) and (border_mode != cv2.BORDER_CONSTANT):
+        img_patch[:,:,3] = cv2.warpAffine(img[:,:,3], trans, (int(patch_width), int(patch_height)), 
+                                            flags=cv2.INTER_LINEAR, 
+                                            borderMode=cv2.BORDER_CONSTANT,
+                            )
+
+    return img_patch, trans
+
+
+def convert_cvimg_to_tensor(cvimg: np.array):
+    """
+    Convert image from HWC to CHW format.
+    Args:
+        cvimg (np.array): Image of shape (H, W, 3) as loaded by OpenCV.
+    Returns:
+        np.array: Output image of shape (3, H, W).
+    """
+    # from h,w,c(OpenCV) to c,h,w
+    img = cvimg.copy()
+    img = np.transpose(img, (2, 0, 1))
+    # from int to float
+    img = img.astype(np.float32)
+    return img
+
+def fliplr_params(smpl_params: Dict, has_smpl_params: Dict) -> Tuple[Dict, Dict]:
+    """
+    Flip SMPL parameters when flipping the image.
+    Args:
+        smpl_params (Dict): SMPL parameter annotations.
+        has_smpl_params (Dict): Whether SMPL annotations are valid.
+    Returns:
+        Dict, Dict: Flipped SMPL parameters and valid flags.
+    """
+    global_orient = smpl_params['global_orient'].copy()
+    body_pose = smpl_params['body_pose'].copy()
+    betas = smpl_params['betas'].copy()
+    has_global_orient = has_smpl_params['global_orient'].copy()
+    has_body_pose = has_smpl_params['body_pose'].copy()
+    has_betas = has_smpl_params['betas'].copy()
+
+    body_pose_permutation = [6, 7, 8, 3, 4, 5, 9, 10, 11, 15, 16, 17, 12, 13,
+                             14 ,18, 19, 20, 24, 25, 26, 21, 22, 23, 27, 28, 29, 33,
+                             34, 35, 30, 31, 32, 36, 37, 38, 42, 43, 44, 39, 40, 41,
+                             45, 46, 47, 51, 52, 53, 48, 49, 50, 57, 58, 59, 54, 55,
+                             56, 63, 64, 65, 60, 61, 62, 69, 70, 71, 66, 67, 68]
+    body_pose_permutation = body_pose_permutation[:len(body_pose)]
+    body_pose_permutation = [i-3 for i in body_pose_permutation]
+
+    body_pose = body_pose[body_pose_permutation]
+
+    global_orient[1::3] *= -1
+    global_orient[2::3] *= -1
+    body_pose[1::3] *= -1
+    body_pose[2::3] *= -1
+
+    smpl_params = {'global_orient': global_orient.astype(np.float32),
+                   'body_pose': body_pose.astype(np.float32),
+                   'betas': betas.astype(np.float32)
+                  }
+
+    has_smpl_params = {'global_orient': has_global_orient,
+                       'body_pose': has_body_pose,
+                       'betas': has_betas
+                      }
+
+    return smpl_params, has_smpl_params
+
+
+def fliplr_keypoints(joints: np.array, width: float, flip_permutation: List[int]) -> np.array:
+    """
+    Flip 2D or 3D keypoints.
+    Args:
+        joints (np.array): Array of shape (N, 3) or (N, 4) containing 2D or 3D keypoint locations and confidence.
+        flip_permutation (List): Permutation to apply after flipping.
+    Returns:
+        np.array: Flipped 2D or 3D keypoints with shape (N, 3) or (N, 4) respectively.
+    """
+    joints = joints.copy()
+    # Flip horizontal
+    joints[:, 0] = width - joints[:, 0] - 1
+    joints = joints[flip_permutation, :]
+
+    return joints
+
+def keypoint_3d_processing(keypoints_3d: np.array, flip_permutation: List[int], rot: float, do_flip: float) -> np.array:
+    """
+    Process 3D keypoints (rotation/flipping).
+    Args:
+        keypoints_3d (np.array): Input array of shape (N, 4) containing the 3D keypoints and confidence.
+        flip_permutation (List): Permutation to apply after flipping.
+        rot (float): Random rotation applied to the keypoints.
+        do_flip (bool): Whether to flip keypoints or not.
+    Returns:
+        np.array: Transformed 3D keypoints with shape (N, 4).
+    """
+    if do_flip:
+        keypoints_3d = fliplr_keypoints(keypoints_3d, 1, flip_permutation)
+    # in-plane rotation
+    rot_mat = np.eye(3)
+    if not rot == 0:
+        rot_rad = -rot * np.pi / 180
+        sn,cs = np.sin(rot_rad), np.cos(rot_rad)
+        rot_mat[0,:2] = [cs, -sn]
+        rot_mat[1,:2] = [sn, cs]
+    keypoints_3d[:, :-1] = np.einsum('ij,kj->ki', rot_mat, keypoints_3d[:, :-1])
+    # flip the x coordinates
+    keypoints_3d = keypoints_3d.astype('float32')
+    return keypoints_3d
+
+def rot_aa(aa: np.array, rot: float) -> np.array:
+    """
+    Rotate axis angle parameters.
+    Args:
+        aa (np.array): Axis-angle vector of shape (3,).
+        rot (np.array): Rotation angle in degrees.
+    Returns:
+        np.array: Rotated axis-angle vector.
+    """
+    # pose parameters
+    R = np.array([[np.cos(np.deg2rad(-rot)), -np.sin(np.deg2rad(-rot)), 0],
+                  [np.sin(np.deg2rad(-rot)), np.cos(np.deg2rad(-rot)), 0],
+                  [0, 0, 1]])
+    # find the rotation of the body in camera frame
+    per_rdg, _ = cv2.Rodrigues(aa)
+    # apply the global rotation to the global orientation
+    resrot, _ = cv2.Rodrigues(np.dot(R,per_rdg))
+    aa = (resrot.T)[0]
+    return aa.astype(np.float32)
+
+def smpl_param_processing(smpl_params: Dict, has_smpl_params: Dict, rot: float, do_flip: bool) -> Tuple[Dict, Dict]:
+    """
+    Apply random augmentations to the SMPL parameters.
+    Args:
+        smpl_params (Dict): SMPL parameter annotations.
+        has_smpl_params (Dict): Whether SMPL annotations are valid.
+        rot (float): Random rotation applied to the keypoints.
+        do_flip (bool): Whether to flip keypoints or not.
+    Returns:
+        Dict, Dict: Transformed SMPL parameters and valid flags.
+    """
+    if do_flip:
+        smpl_params, has_smpl_params = fliplr_params(smpl_params, has_smpl_params)
+    smpl_params['global_orient'] = rot_aa(smpl_params['global_orient'], rot)
+    return smpl_params, has_smpl_params
+
+
+
+def get_example(img_path: str|np.ndarray, center_x: float, center_y: float,
+                width: float, height: float,
+                keypoints_2d: np.array, keypoints_3d: np.array,
+                smpl_params: Dict, has_smpl_params: Dict,
+                flip_kp_permutation: List[int],
+                patch_width: int, patch_height: int,
+                mean: np.array, std: np.array,
+                do_augment: bool, augm_config: CfgNode,
+                is_bgr: bool = True,
+                use_skimage_antialias: bool = False,
+                border_mode: int = cv2.BORDER_CONSTANT,
+                return_trans: bool = False) -> Tuple:
+    """
+    Get an example from the dataset and (possibly) apply random augmentations.
+    Args:
+        img_path (str): Image filename
+        center_x (float): Bounding box center x coordinate in the original image.
+        center_y (float): Bounding box center y coordinate in the original image.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array with shape (N,3) containing the 2D keypoints in the original image coordinates.
+        keypoints_3d (np.array): Array with shape (N,4) containing the 3D keypoints.
+        smpl_params (Dict): SMPL parameter annotations.
+        has_smpl_params (Dict): Whether SMPL annotations are valid.
+        flip_kp_permutation (List): Permutation to apply to the keypoints after flipping.
+        patch_width (float): Output box width.
+        patch_height (float): Output box height.
+        mean (np.array): Array of shape (3,) containing the mean for normalizing the input image.
+        std (np.array): Array of shape (3,) containing the std for normalizing the input image.
+        do_augment (bool): Whether to apply data augmentation or not.
+        aug_config (CfgNode): Config containing augmentation parameters.
+    Returns:
+        return img_patch, keypoints_2d, keypoints_3d, smpl_params, has_smpl_params, img_size
+        img_patch (np.array): Cropped image patch of shape (3, patch_height, patch_height)
+        keypoints_2d (np.array): Array with shape (N,3) containing the transformed 2D keypoints.
+        keypoints_3d (np.array): Array with shape (N,4) containing the transformed 3D keypoints.
+        smpl_params (Dict): Transformed SMPL parameters.
+        has_smpl_params (Dict): Valid flag for transformed SMPL parameters.
+        img_size (np.array): Image size of the original image.
+        """
+    if isinstance(img_path, str):
+        # 1. load image
+        cvimg = cv2.imread(img_path, cv2.IMREAD_COLOR | cv2.IMREAD_IGNORE_ORIENTATION)
+        if not isinstance(cvimg, np.ndarray):
+            raise IOError("Fail to read %s" % img_path)
+    elif isinstance(img_path, np.ndarray):
+        cvimg = img_path
+    else:
+        raise TypeError('img_path must be either a string or a numpy array')
+    img_height, img_width, img_channels = cvimg.shape
+
+    img_size = np.array([img_height, img_width])
+
+    # 2. get augmentation params
+    if do_augment:
+        scale, rot, do_flip, do_extreme_crop, extreme_crop_lvl, color_scale, tx, ty = do_augmentation(augm_config)
+    else:
+        scale, rot, do_flip, do_extreme_crop, extreme_crop_lvl, color_scale, tx, ty = 1.0, 0, False, False, 0, [1.0, 1.0, 1.0], 0., 0.
+
+    if width < 1 or height < 1:
+        breakpoint()
+
+    if do_extreme_crop:
+        if extreme_crop_lvl == 0:
+            center_x1, center_y1, width1, height1 = extreme_cropping(center_x, center_y, width, height, keypoints_2d)
+        elif extreme_crop_lvl == 1:
+            center_x1, center_y1, width1, height1 = extreme_cropping_aggressive(center_x, center_y, width, height, keypoints_2d)
+
+        THRESH = 4
+        if width1 < THRESH or height1 < THRESH:
+            # print(f'{do_extreme_crop=}')
+            # print(f'width: {width}, height: {height}')
+            # print(f'width1: {width1}, height1: {height1}')
+            # print(f'center_x: {center_x}, center_y: {center_y}')
+            # print(f'center_x1: {center_x1}, center_y1: {center_y1}')
+            # print(f'keypoints_2d: {keypoints_2d}')
+            # print(f'\n\n', flush=True)
+            # breakpoint()
+            pass
+            # print(f'skip ==> width1: {width1}, height1: {height1}, width: {width}, height: {height}')
+        else:
+            center_x, center_y, width, height = center_x1, center_y1, width1, height1
+
+    center_x += width * tx
+    center_y += height * ty
+
+    # Process 3D keypoints
+    keypoints_3d = keypoint_3d_processing(keypoints_3d, flip_kp_permutation, rot, do_flip)
+
+    # 3. generate image patch
+    if use_skimage_antialias:
+        # Blur image to avoid aliasing artifacts
+        downsampling_factor = (patch_width / (width*scale))
+        if downsampling_factor > 1.1:
+            cvimg  = gaussian(cvimg, sigma=(downsampling_factor-1)/2, channel_axis=2, preserve_range=True, truncate=3.0)
+
+    img_patch_cv, trans = generate_image_patch_cv2(cvimg,
+                                                    center_x, center_y,
+                                                    width, height,
+                                                    patch_width, patch_height,
+                                                    do_flip, scale, rot, 
+                                                    border_mode=border_mode)
+        # img_patch_cv, trans = generate_image_patch_skimage(cvimg,
+        #                                                 center_x, center_y,
+        #                                                 width, height,
+        #                                                 patch_width, patch_height,
+        #                                                 do_flip, scale, rot, 
+        #                                                 border_mode=border_mode)
+
+    image = img_patch_cv.copy()
+    if is_bgr:
+        image = image[:, :, ::-1]
+    img_patch_cv = image.copy()
+    img_patch = convert_cvimg_to_tensor(image)
+
+
+    smpl_params, has_smpl_params = smpl_param_processing(smpl_params, has_smpl_params, rot, do_flip)
+
+    # apply normalization
+    for n_c in range(min(img_channels, 3)):
+        img_patch[n_c, :, :] = np.clip(img_patch[n_c, :, :] * color_scale[n_c], 0, 255)
+        if mean is not None and std is not None:
+            img_patch[n_c, :, :] = (img_patch[n_c, :, :] - mean[n_c]) / std[n_c]
+    if do_flip:
+        keypoints_2d = fliplr_keypoints(keypoints_2d, img_width, flip_kp_permutation)
+
+
+    for n_jt in range(len(keypoints_2d)):
+        keypoints_2d[n_jt, 0:2] = trans_point2d(keypoints_2d[n_jt, 0:2], trans)
+    keypoints_2d[:, :-1] = keypoints_2d[:, :-1] / patch_width - 0.5
+
+    if not return_trans:
+        return img_patch, keypoints_2d, keypoints_3d, smpl_params, has_smpl_params, img_size
+    else:
+        return img_patch, keypoints_2d, keypoints_3d, smpl_params, has_smpl_params, img_size, trans
+
+def crop_to_hips(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array) -> Tuple:
+    """
+    Extreme cropping: Crop the box up to the hip locations.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    lower_body_keypoints = [10, 11, 13, 14, 19, 20, 21, 22, 23, 24, 25+0, 25+1, 25+4, 25+5]
+    keypoints_2d[lower_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+
+def crop_to_shoulders(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box up to the shoulder locations.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    lower_body_keypoints = [3, 4, 6, 7, 8, 9, 10, 11, 12, 13, 14, 19, 20, 21, 22, 23, 24] + [25 + i for i in [0, 1, 2, 3, 4, 5, 6, 7, 10, 11, 14, 15, 16]]
+    keypoints_2d[lower_body_keypoints, :] = 0
+    center, scale = get_bbox(keypoints_2d)
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.2 * scale[0]
+        height = 1.2 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_to_head(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the head.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    lower_body_keypoints = [3, 4, 6, 7, 8, 9, 10, 11, 12, 13, 14, 19, 20, 21, 22, 23, 24] + [25 + i for i in [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 14, 15, 16]]
+    keypoints_2d[lower_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.3 * scale[0]
+        height = 1.3 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_torso_only(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the torso.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    nontorso_body_keypoints = [0, 3, 4, 6, 7, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24] + [25 + i for i in [0, 1, 4, 5, 6, 7, 10, 11, 13, 17, 18]]
+    keypoints_2d[nontorso_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_rightarm_only(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the right arm.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    nonrightarm_body_keypoints = [0, 1, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24] + [25 + i for i in [0, 1, 2, 3, 4, 5, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18]]
+    keypoints_2d[nonrightarm_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_leftarm_only(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the left arm.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    nonleftarm_body_keypoints = [0, 1, 2, 3, 4, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24] + [25 + i for i in [0, 1, 2, 3, 4, 5, 6, 7, 8, 12, 13, 14, 15, 16, 17, 18]]
+    keypoints_2d[nonleftarm_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_legs_only(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the legs.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    nonlegs_body_keypoints = [0, 1, 2, 3, 4, 5, 6, 7, 15, 16, 17, 18] + [25 + i for i in [6, 7, 8, 9, 10, 11, 12, 13, 15, 16, 17, 18]]
+    keypoints_2d[nonlegs_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_rightleg_only(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the right leg.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    nonrightleg_body_keypoints = [0, 1, 2, 3, 4, 5, 6, 7, 8, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21] + [25 + i for i in [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18]]
+    keypoints_2d[nonrightleg_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+def crop_leftleg_only(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array):
+    """
+    Extreme cropping: Crop the box and keep on only the left leg.
+    Args:
+        center_x (float): x coordinate of the bounding box center.
+        center_y (float): y coordinate of the bounding box center.
+        width (float): Bounding box width.
+        height (float): Bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        center_x (float): x coordinate of the new bounding box center.
+        center_y (float): y coordinate of the new bounding box center.
+        width (float): New bounding box width.
+        height (float): New bounding box height.
+    """
+    keypoints_2d = keypoints_2d.copy()
+    nonleftleg_body_keypoints = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 15, 16, 17, 18, 22, 23, 24] + [25 + i for i in [0, 1, 2, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18]]
+    keypoints_2d[nonleftleg_body_keypoints, :] = 0
+    if keypoints_2d[:, -1].sum() > 1:
+        center, scale = get_bbox(keypoints_2d)
+        center_x = center[0]
+        center_y = center[1]
+        width = 1.1 * scale[0]
+        height = 1.1 * scale[1]
+    return center_x, center_y, width, height
+
+def full_body(keypoints_2d: np.array) -> bool:
+    """
+    Check if all main body joints are visible.
+    Args:
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        bool: True if all main body joints are visible.
+    """
+
+    body_keypoints_openpose = [2, 3, 4, 5, 6, 7, 10, 11, 13, 14]
+    body_keypoints = [25 + i for i in [8, 7, 6, 9, 10, 11, 1, 0, 4, 5]]
+    return (np.maximum(keypoints_2d[body_keypoints, -1], keypoints_2d[body_keypoints_openpose, -1]) > 0).sum() == len(body_keypoints)
+
+def upper_body(keypoints_2d: np.array):
+    """
+    Check if all upper body joints are visible.
+    Args:
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+    Returns:
+        bool: True if all main body joints are visible.
+    """
+    lower_body_keypoints_openpose = [10, 11, 13, 14]
+    lower_body_keypoints = [25 + i for i in [1, 0, 4, 5]]
+    upper_body_keypoints_openpose = [0, 1, 15, 16, 17, 18]
+    upper_body_keypoints = [25+8, 25+9, 25+12, 25+13, 25+17, 25+18]
+    return ((keypoints_2d[lower_body_keypoints + lower_body_keypoints_openpose, -1] > 0).sum() == 0)\
+       and ((keypoints_2d[upper_body_keypoints + upper_body_keypoints_openpose, -1] > 0).sum() >= 2)
+
+def get_bbox(keypoints_2d: np.array, rescale: float = 1.2) -> Tuple:
+    """
+    Get center and scale for bounding box from openpose detections.
+    Args:
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+        rescale (float): Scale factor to rescale bounding boxes computed from the keypoints.
+    Returns:
+        center (np.array): Array of shape (2,) containing the new bounding box center.
+        scale (float): New bounding box scale.
+    """
+    valid = keypoints_2d[:,-1] > 0
+    valid_keypoints = keypoints_2d[valid][:,:-1]
+    center = 0.5 * (valid_keypoints.max(axis=0) + valid_keypoints.min(axis=0))
+    bbox_size = (valid_keypoints.max(axis=0) - valid_keypoints.min(axis=0))
+    # adjust bounding box tightness
+    scale = bbox_size
+    scale *= rescale
+    return center, scale
+
+def extreme_cropping(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array) -> Tuple:
+    """
+    Perform extreme cropping
+    Args:
+        center_x (float): x coordinate of bounding box center.
+        center_y (float): y coordinate of bounding box center.
+        width (float): bounding box width.
+        height (float): bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+        rescale (float): Scale factor to rescale bounding boxes computed from the keypoints.
+    Returns:
+        center_x (float): x coordinate of bounding box center.
+        center_y (float): y coordinate of bounding box center.
+        width (float): bounding box width.
+        height (float): bounding box height.
+    """
+    p = torch.rand(1).item()
+    if full_body(keypoints_2d):
+        if p < 0.7:
+            center_x, center_y, width, height = crop_to_hips(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.9:
+            center_x, center_y, width, height = crop_to_shoulders(center_x, center_y, width, height, keypoints_2d)
+        else:
+            center_x, center_y, width, height = crop_to_head(center_x, center_y, width, height, keypoints_2d)
+    elif upper_body(keypoints_2d):
+        if p < 0.9:
+            center_x, center_y, width, height = crop_to_shoulders(center_x, center_y, width, height, keypoints_2d)
+        else:
+            center_x, center_y, width, height = crop_to_head(center_x, center_y, width, height, keypoints_2d)
+
+    return center_x, center_y, max(width, height), max(width, height)
+
+def extreme_cropping_aggressive(center_x: float, center_y: float, width: float, height: float, keypoints_2d: np.array) -> Tuple:
+    """
+    Perform aggressive extreme cropping
+    Args:
+        center_x (float): x coordinate of bounding box center.
+        center_y (float): y coordinate of bounding box center.
+        width (float): bounding box width.
+        height (float): bounding box height.
+        keypoints_2d (np.array): Array of shape (N, 3) containing 2D keypoint locations.
+        rescale (float): Scale factor to rescale bounding boxes computed from the keypoints.
+    Returns:
+        center_x (float): x coordinate of bounding box center.
+        center_y (float): y coordinate of bounding box center.
+        width (float): bounding box width.
+        height (float): bounding box height.
+    """
+    p = torch.rand(1).item()
+    if full_body(keypoints_2d):
+        if p < 0.2:
+            center_x, center_y, width, height = crop_to_hips(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.3:
+            center_x, center_y, width, height = crop_to_shoulders(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.4:
+            center_x, center_y, width, height = crop_to_head(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.5:
+            center_x, center_y, width, height = crop_torso_only(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.6:
+            center_x, center_y, width, height = crop_rightarm_only(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.7:
+            center_x, center_y, width, height = crop_leftarm_only(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.8:
+            center_x, center_y, width, height = crop_legs_only(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.9:
+            center_x, center_y, width, height = crop_rightleg_only(center_x, center_y, width, height, keypoints_2d)
+        else:
+            center_x, center_y, width, height = crop_leftleg_only(center_x, center_y, width, height, keypoints_2d)
+    elif upper_body(keypoints_2d):
+        if p < 0.2:
+            center_x, center_y, width, height = crop_to_shoulders(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.4:
+            center_x, center_y, width, height = crop_to_head(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.6:
+            center_x, center_y, width, height = crop_torso_only(center_x, center_y, width, height, keypoints_2d)
+        elif p < 0.8:
+            center_x, center_y, width, height = crop_rightarm_only(center_x, center_y, width, height, keypoints_2d)
+        else:
+            center_x, center_y, width, height = crop_leftarm_only(center_x, center_y, width, height, keypoints_2d)
+    return center_x, center_y, max(width, height), max(width, height)

hmr2/datasets/vitdet_dataset.py

@@ -0,0 +1,89 @@
+from typing import Dict
+
+import cv2
+import numpy as np
+from skimage.filters import gaussian
+from yacs.config import CfgNode
+import torch
+
+from .utils import (convert_cvimg_to_tensor,
+                    expand_to_aspect_ratio,
+                    generate_image_patch_cv2)
+
+DEFAULT_MEAN = 255. * np.array([0.485, 0.456, 0.406])
+DEFAULT_STD = 255. * np.array([0.229, 0.224, 0.225])
+
+class ViTDetDataset(torch.utils.data.Dataset):
+
+    def __init__(self,
+                 cfg: CfgNode,
+                 img_cv2: np.array,
+                 boxes: np.array,
+                 train: bool = False,
+                 **kwargs):
+        super().__init__()
+        self.cfg = cfg
+        self.img_cv2 = img_cv2
+        # self.boxes = boxes
+
+        assert train == False, "ViTDetDataset is only for inference"
+        self.train = train
+        self.img_size = cfg.MODEL.IMAGE_SIZE
+        self.mean = 255. * np.array(self.cfg.MODEL.IMAGE_MEAN)
+        self.std = 255. * np.array(self.cfg.MODEL.IMAGE_STD)
+
+        # Preprocess annotations
+        boxes = boxes.astype(np.float32)
+        self.center = (boxes[:, 2:4] + boxes[:, 0:2]) / 2.0
+        self.scale = (boxes[:, 2:4] - boxes[:, 0:2]) / 200.0
+        self.personid = np.arange(len(boxes), dtype=np.int32)
+
+    def __len__(self) -> int:
+        return len(self.personid)
+    
+    def __getitem__(self, idx: int) -> Dict[str, np.array]:
+
+        center = self.center[idx].copy()
+        center_x = center[0]
+        center_y = center[1]
+
+        scale = self.scale[idx]
+        BBOX_SHAPE = self.cfg.MODEL.get('BBOX_SHAPE', None)
+        bbox_size = expand_to_aspect_ratio(scale*200, target_aspect_ratio=BBOX_SHAPE).max()
+
+        patch_width = patch_height = self.img_size
+
+        # 3. generate image patch
+        # if use_skimage_antialias:
+        cvimg = self.img_cv2.copy()
+        if True:
+            # Blur image to avoid aliasing artifacts
+            downsampling_factor = ((bbox_size*1.0) / patch_width)
+            print(f'{downsampling_factor=}')
+            downsampling_factor = downsampling_factor / 2.0
+            if downsampling_factor > 1.1:
+                cvimg  = gaussian(cvimg, sigma=(downsampling_factor-1)/2, channel_axis=2, preserve_range=True)
+
+
+        img_patch_cv, trans = generate_image_patch_cv2(cvimg,
+                                                    center_x, center_y,
+                                                    bbox_size, bbox_size,
+                                                    patch_width, patch_height,
+                                                    False, 1.0, 0, 
+                                                    border_mode=cv2.BORDER_CONSTANT)
+        img_patch_cv = img_patch_cv[:, :, ::-1]
+        img_patch = convert_cvimg_to_tensor(img_patch_cv)
+
+        # apply normalization
+        for n_c in range(min(self.img_cv2.shape[2], 3)):
+            img_patch[n_c, :, :] = (img_patch[n_c, :, :] - self.mean[n_c]) / self.std[n_c]
+
+
+        item = {
+            'img': img_patch,
+            'personid': int(self.personid[idx]),
+        }        
+        item['box_center'] = self.center[idx].copy()
+        item['box_size'] = bbox_size
+        item['img_size'] = 1.0 * np.array([cvimg.shape[1], cvimg.shape[0]])
+        return item

hmr2/models/__init__.py

@@ -0,0 +1,3 @@
+from .smpl_wrapper import SMPL
+from .hmr2 import HMR2
+from .discriminator import Discriminator

hmr2/models/backbones/__init__.py

@@ -0,0 +1,7 @@
+from .vit import vit
+
+def create_backbone(cfg):
+    if cfg.MODEL.BACKBONE.TYPE == 'vit':
+        return vit(cfg)
+    else:
+        raise NotImplementedError('Backbone type is not implemented')

hmr2/models/backbones/vit.py

@@ -0,0 +1,348 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import math
+
+import torch
+from functools import partial
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+
+from timm.models.layers import drop_path, to_2tuple, trunc_normal_
+
+def vit(cfg):
+    return ViT(
+                img_size=(256, 192),
+                patch_size=16,
+                embed_dim=1280,
+                depth=32,
+                num_heads=16,
+                ratio=1,
+                use_checkpoint=False,
+                mlp_ratio=4,
+                qkv_bias=True,
+                drop_path_rate=0.55,
+            )
+
+def get_abs_pos(abs_pos, h, w, ori_h, ori_w, has_cls_token=True):
+    """
+    Calculate absolute positional embeddings. If needed, resize embeddings and remove cls_token
+        dimension for the original embeddings.
+    Args:
+        abs_pos (Tensor): absolute positional embeddings with (1, num_position, C).
+        has_cls_token (bool): If true, has 1 embedding in abs_pos for cls token.
+        hw (Tuple): size of input image tokens.
+
+    Returns:
+        Absolute positional embeddings after processing with shape (1, H, W, C)
+    """
+    cls_token = None
+    B, L, C = abs_pos.shape
+    if has_cls_token:
+        cls_token = abs_pos[:, 0:1]
+        abs_pos = abs_pos[:, 1:]
+
+    if ori_h != h or ori_w != w:
+        new_abs_pos = F.interpolate(
+            abs_pos.reshape(1, ori_h, ori_w, -1).permute(0, 3, 1, 2),
+            size=(h, w),
+            mode="bicubic",
+            align_corners=False,
+        ).permute(0, 2, 3, 1).reshape(B, -1, C)
+
+    else:
+        new_abs_pos = abs_pos
+    
+    if cls_token is not None:
+        new_abs_pos = torch.cat([cls_token, new_abs_pos], dim=1)
+    return new_abs_pos
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+    
+    def extra_repr(self):
+        return 'p={}'.format(self.drop_prob)
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+class Attention(nn.Module):
+    def __init__(
+            self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,
+            proj_drop=0., attn_head_dim=None,):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.dim = dim
+
+        if attn_head_dim is not None:
+            head_dim = attn_head_dim
+        all_head_dim = head_dim * self.num_heads
+
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.qkv = nn.Linear(dim, all_head_dim * 3, bias=qkv_bias)
+
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(all_head_dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv = self.qkv(x)
+        qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+
+        return x
+
+class Block(nn.Module):
+
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, 
+                 drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, 
+                 norm_layer=nn.LayerNorm, attn_head_dim=None
+                 ):
+        super().__init__()
+        
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, attn_head_dim=attn_head_dim
+            )
+
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x):
+        x = x + self.drop_path(self.attn(self.norm1(x)))
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, ratio=1):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0]) * (ratio ** 2)
+        self.patch_shape = (int(img_size[0] // patch_size[0] * ratio), int(img_size[1] // patch_size[1] * ratio))
+        self.origin_patch_shape = (int(img_size[0] // patch_size[0]), int(img_size[1] // patch_size[1]))
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.num_patches = num_patches
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=(patch_size[0] // ratio), padding=4 + 2 * (ratio//2-1))
+
+    def forward(self, x, **kwargs):
+        B, C, H, W = x.shape
+        x = self.proj(x)
+        Hp, Wp = x.shape[2], x.shape[3]
+
+        x = x.flatten(2).transpose(1, 2)
+        return x, (Hp, Wp)
+
+
+class HybridEmbed(nn.Module):
+    """ CNN Feature Map Embedding
+    Extract feature map from CNN, flatten, project to embedding dim.
+    """
+    def __init__(self, backbone, img_size=224, feature_size=None, in_chans=3, embed_dim=768):
+        super().__init__()
+        assert isinstance(backbone, nn.Module)
+        img_size = to_2tuple(img_size)
+        self.img_size = img_size
+        self.backbone = backbone
+        if feature_size is None:
+            with torch.no_grad():
+                training = backbone.training
+                if training:
+                    backbone.eval()
+                o = self.backbone(torch.zeros(1, in_chans, img_size[0], img_size[1]))[-1]
+                feature_size = o.shape[-2:]
+                feature_dim = o.shape[1]
+                backbone.train(training)
+        else:
+            feature_size = to_2tuple(feature_size)
+            feature_dim = self.backbone.feature_info.channels()[-1]
+        self.num_patches = feature_size[0] * feature_size[1]
+        self.proj = nn.Linear(feature_dim, embed_dim)
+
+    def forward(self, x):
+        x = self.backbone(x)[-1]
+        x = x.flatten(2).transpose(1, 2)
+        x = self.proj(x)
+        return x
+
+
+class ViT(nn.Module):
+
+    def __init__(self,
+                 img_size=224, patch_size=16, in_chans=3, num_classes=80, embed_dim=768, depth=12,
+                 num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
+                 drop_path_rate=0., hybrid_backbone=None, norm_layer=None, use_checkpoint=False, 
+                 frozen_stages=-1, ratio=1, last_norm=True,
+                 patch_padding='pad', freeze_attn=False, freeze_ffn=False,
+                 ):
+        # Protect mutable default arguments
+        super(ViT, self).__init__()
+        norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)
+        self.num_classes = num_classes
+        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+        self.frozen_stages = frozen_stages
+        self.use_checkpoint = use_checkpoint
+        self.patch_padding = patch_padding
+        self.freeze_attn = freeze_attn
+        self.freeze_ffn = freeze_ffn
+        self.depth = depth
+
+        if hybrid_backbone is not None:
+            self.patch_embed = HybridEmbed(
+                hybrid_backbone, img_size=img_size, in_chans=in_chans, embed_dim=embed_dim)
+        else:
+            self.patch_embed = PatchEmbed(
+                img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, ratio=ratio)
+        num_patches = self.patch_embed.num_patches
+
+        # since the pretraining model has class token
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
+
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
+
+        self.blocks = nn.ModuleList([
+            Block(
+                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
+                )
+            for i in range(depth)])
+
+        self.last_norm = norm_layer(embed_dim) if last_norm else nn.Identity()
+
+        if self.pos_embed is not None:
+            trunc_normal_(self.pos_embed, std=.02)
+
+        self._freeze_stages()
+
+    def _freeze_stages(self):
+        """Freeze parameters."""
+        if self.frozen_stages >= 0:
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+
+        for i in range(1, self.frozen_stages + 1):
+            m = self.blocks[i]
+            m.eval()
+            for param in m.parameters():
+                param.requires_grad = False
+
+        if self.freeze_attn:
+            for i in range(0, self.depth):
+                m = self.blocks[i]
+                m.attn.eval()
+                m.norm1.eval()
+                for param in m.attn.parameters():
+                    param.requires_grad = False
+                for param in m.norm1.parameters():
+                    param.requires_grad = False
+
+        if self.freeze_ffn:
+            self.pos_embed.requires_grad = False
+            self.patch_embed.eval()
+            for param in self.patch_embed.parameters():
+                param.requires_grad = False
+            for i in range(0, self.depth):
+                m = self.blocks[i]
+                m.mlp.eval()
+                m.norm2.eval()
+                for param in m.mlp.parameters():
+                    param.requires_grad = False
+                for param in m.norm2.parameters():
+                    param.requires_grad = False
+
+    def init_weights(self):
+        """Initialize the weights in backbone.
+        Args:
+            pretrained (str, optional): Path to pre-trained weights.
+                Defaults to None.
+        """
+        def _init_weights(m):
+            if isinstance(m, nn.Linear):
+                trunc_normal_(m.weight, std=.02)
+                if isinstance(m, nn.Linear) and m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.LayerNorm):
+                nn.init.constant_(m.bias, 0)
+                nn.init.constant_(m.weight, 1.0)
+
+        self.apply(_init_weights)
+
+    def get_num_layers(self):
+        return len(self.blocks)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed', 'cls_token'}
+
+    def forward_features(self, x):
+        B, C, H, W = x.shape
+        x, (Hp, Wp) = self.patch_embed(x)
+
+        if self.pos_embed is not None:
+            # fit for multiple GPU training
+            # since the first element for pos embed (sin-cos manner) is zero, it will cause no difference
+            x = x + self.pos_embed[:, 1:] + self.pos_embed[:, :1]
+
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x = blk(x)
+
+        x = self.last_norm(x)
+
+        xp = x.permute(0, 2, 1).reshape(B, -1, Hp, Wp).contiguous()
+
+        return xp
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        return x
+
+    def train(self, mode=True):
+        """Convert the model into training mode."""
+        super().train(mode)
+        self._freeze_stages()

hmr2/models/backbones/vit_vitpose.py

@@ -0,0 +1,17 @@
+# import mmcv
+# import mmpose
+# from mmpose.models import build_posenet
+# from mmcv.runner import load_checkpoint
+# from pathlib import Path
+
+# def vit(cfg):
+#     vitpose_dir = Path(mmpose.__file__).parent.parent
+#     config = f'{vitpose_dir}/configs/body/2d_kpt_sview_rgb_img/topdown_heatmap/coco/ViTPose_huge_coco_256x192.py'
+#     # checkpoint = f'{vitpose_dir}/models/vitpose-h-multi-coco.pth'
+
+#     config = mmcv.Config.fromfile(config)
+#     config.model.pretrained = None
+#     model = build_posenet(config.model)
+#     # load_checkpoint(model, checkpoint, map_location='cpu')
+
+#     return model.backbone

hmr2/models/components/pose_transformer.py

@@ -0,0 +1,358 @@
+from inspect import isfunction
+from typing import Callable, Optional
+
+import torch
+from einops import rearrange
+from einops.layers.torch import Rearrange
+from torch import nn
+
+from .t_cond_mlp import (
+    AdaptiveLayerNorm1D,
+    FrequencyEmbedder,
+    normalization_layer,
+)
+# from .vit import Attention, FeedForward
+
+
+def exists(val):
+    return val is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+class PreNorm(nn.Module):
+    def __init__(self, dim: int, fn: Callable, norm: str = "layer", norm_cond_dim: int = -1):
+        super().__init__()
+        self.norm = normalization_layer(norm, dim, norm_cond_dim)
+        self.fn = fn
+
+    def forward(self, x: torch.Tensor, *args, **kwargs):
+        if isinstance(self.norm, AdaptiveLayerNorm1D):
+            return self.fn(self.norm(x, *args), **kwargs)
+        else:
+            return self.fn(self.norm(x), **kwargs)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, hidden_dim, dropout=0.0):
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(dim, hidden_dim),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(hidden_dim, dim),
+            nn.Dropout(dropout),
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, heads=8, dim_head=64, dropout=0.0):
+        super().__init__()
+        inner_dim = dim_head * heads
+        project_out = not (heads == 1 and dim_head == dim)
+
+        self.heads = heads
+        self.scale = dim_head**-0.5
+
+        self.attend = nn.Softmax(dim=-1)
+        self.dropout = nn.Dropout(dropout)
+
+        self.to_qkv = nn.Linear(dim, inner_dim * 3, bias=False)
+
+        self.to_out = (
+            nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout))
+            if project_out
+            else nn.Identity()
+        )
+
+    def forward(self, x):
+        qkv = self.to_qkv(x).chunk(3, dim=-1)
+        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=self.heads), qkv)
+
+        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
+
+        attn = self.attend(dots)
+        attn = self.dropout(attn)
+
+        out = torch.matmul(attn, v)
+        out = rearrange(out, "b h n d -> b n (h d)")
+        return self.to_out(out)
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, dim, context_dim=None, heads=8, dim_head=64, dropout=0.0):
+        super().__init__()
+        inner_dim = dim_head * heads
+        project_out = not (heads == 1 and dim_head == dim)
+
+        self.heads = heads
+        self.scale = dim_head**-0.5
+
+        self.attend = nn.Softmax(dim=-1)
+        self.dropout = nn.Dropout(dropout)
+
+        context_dim = default(context_dim, dim)
+        self.to_kv = nn.Linear(context_dim, inner_dim * 2, bias=False)
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+
+        self.to_out = (
+            nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout))
+            if project_out
+            else nn.Identity()
+        )
+
+    def forward(self, x, context=None):
+        context = default(context, x)
+        k, v = self.to_kv(context).chunk(2, dim=-1)
+        q = self.to_q(x)
+        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=self.heads), [q, k, v])
+
+        dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
+
+        attn = self.attend(dots)
+        attn = self.dropout(attn)
+
+        out = torch.matmul(attn, v)
+        out = rearrange(out, "b h n d -> b n (h d)")
+        return self.to_out(out)
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        depth: int,
+        heads: int,
+        dim_head: int,
+        mlp_dim: int,
+        dropout: float = 0.0,
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+    ):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            sa = Attention(dim, heads=heads, dim_head=dim_head, dropout=dropout)
+            ff = FeedForward(dim, mlp_dim, dropout=dropout)
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PreNorm(dim, sa, norm=norm, norm_cond_dim=norm_cond_dim),
+                        PreNorm(dim, ff, norm=norm, norm_cond_dim=norm_cond_dim),
+                    ]
+                )
+            )
+
+    def forward(self, x: torch.Tensor, *args):
+        for attn, ff in self.layers:
+            x = attn(x, *args) + x
+            x = ff(x, *args) + x
+        return x
+
+
+class TransformerCrossAttn(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        depth: int,
+        heads: int,
+        dim_head: int,
+        mlp_dim: int,
+        dropout: float = 0.0,
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+        context_dim: Optional[int] = None,
+    ):
+        super().__init__()
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            sa = Attention(dim, heads=heads, dim_head=dim_head, dropout=dropout)
+            ca = CrossAttention(
+                dim, context_dim=context_dim, heads=heads, dim_head=dim_head, dropout=dropout
+            )
+            ff = FeedForward(dim, mlp_dim, dropout=dropout)
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PreNorm(dim, sa, norm=norm, norm_cond_dim=norm_cond_dim),
+                        PreNorm(dim, ca, norm=norm, norm_cond_dim=norm_cond_dim),
+                        PreNorm(dim, ff, norm=norm, norm_cond_dim=norm_cond_dim),
+                    ]
+                )
+            )
+
+    def forward(self, x: torch.Tensor, *args, context=None, context_list=None):
+        if context_list is None:
+            context_list = [context] * len(self.layers)
+        if len(context_list) != len(self.layers):
+            raise ValueError(f"len(context_list) != len(self.layers) ({len(context_list)} != {len(self.layers)})")
+
+        for i, (self_attn, cross_attn, ff) in enumerate(self.layers):
+            x = self_attn(x, *args) + x
+            x = cross_attn(x, *args, context=context_list[i]) + x
+            x = ff(x, *args) + x
+        return x
+
+
+class DropTokenDropout(nn.Module):
+    def __init__(self, p: float = 0.1):
+        super().__init__()
+        if p < 0 or p > 1:
+            raise ValueError(
+                "dropout probability has to be between 0 and 1, " "but got {}".format(p)
+            )
+        self.p = p
+
+    def forward(self, x: torch.Tensor):
+        # x: (batch_size, seq_len, dim)
+        if self.training and self.p > 0:
+            zero_mask = torch.full_like(x[0, :, 0], self.p).bernoulli().bool()
+            # TODO: permutation idx for each batch using torch.argsort
+            if zero_mask.any():
+                x = x[:, ~zero_mask, :]
+        return x
+
+
+class ZeroTokenDropout(nn.Module):
+    def __init__(self, p: float = 0.1):
+        super().__init__()
+        if p < 0 or p > 1:
+            raise ValueError(
+                "dropout probability has to be between 0 and 1, " "but got {}".format(p)
+            )
+        self.p = p
+
+    def forward(self, x: torch.Tensor):
+        # x: (batch_size, seq_len, dim)
+        if self.training and self.p > 0:
+            zero_mask = torch.full_like(x[:, :, 0], self.p).bernoulli().bool()
+            # Zero-out the masked tokens
+            x[zero_mask, :] = 0
+        return x
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(
+        self,
+        num_tokens: int,
+        token_dim: int,
+        dim: int,
+        depth: int,
+        heads: int,
+        mlp_dim: int,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        emb_dropout: float = 0.0,
+        emb_dropout_type: str = "drop",
+        emb_dropout_loc: str = "token",
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+        token_pe_numfreq: int = -1,
+    ):
+        super().__init__()
+        if token_pe_numfreq > 0:
+            token_dim_new = token_dim * (2 * token_pe_numfreq + 1)
+            self.to_token_embedding = nn.Sequential(
+                Rearrange("b n d -> (b n) d", n=num_tokens, d=token_dim),
+                FrequencyEmbedder(token_pe_numfreq, token_pe_numfreq - 1),
+                Rearrange("(b n) d -> b n d", n=num_tokens, d=token_dim_new),
+                nn.Linear(token_dim_new, dim),
+            )
+        else:
+            self.to_token_embedding = nn.Linear(token_dim, dim)
+        self.pos_embedding = nn.Parameter(torch.randn(1, num_tokens, dim))
+        if emb_dropout_type == "drop":
+            self.dropout = DropTokenDropout(emb_dropout)
+        elif emb_dropout_type == "zero":
+            self.dropout = ZeroTokenDropout(emb_dropout)
+        else:
+            raise ValueError(f"Unknown emb_dropout_type: {emb_dropout_type}")
+        self.emb_dropout_loc = emb_dropout_loc
+
+        self.transformer = Transformer(
+            dim, depth, heads, dim_head, mlp_dim, dropout, norm=norm, norm_cond_dim=norm_cond_dim
+        )
+
+    def forward(self, inp: torch.Tensor, *args, **kwargs):
+        x = inp
+
+        if self.emb_dropout_loc == "input":
+            x = self.dropout(x)
+        x = self.to_token_embedding(x)
+
+        if self.emb_dropout_loc == "token":
+            x = self.dropout(x)
+        b, n, _ = x.shape
+        x += self.pos_embedding[:, :n]
+
+        if self.emb_dropout_loc == "token_afterpos":
+            x = self.dropout(x)
+        x = self.transformer(x, *args)
+        return x
+
+
+class TransformerDecoder(nn.Module):
+    def __init__(
+        self,
+        num_tokens: int,
+        token_dim: int,
+        dim: int,
+        depth: int,
+        heads: int,
+        mlp_dim: int,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        emb_dropout: float = 0.0,
+        emb_dropout_type: str = 'drop',
+        norm: str = "layer",
+        norm_cond_dim: int = -1,
+        context_dim: Optional[int] = None,
+        skip_token_embedding: bool = False,
+    ):
+        super().__init__()
+        if not skip_token_embedding:
+            self.to_token_embedding = nn.Linear(token_dim, dim)
+        else:
+            self.to_token_embedding = nn.Identity()
+            if token_dim != dim:
+                raise ValueError(
+                    f"token_dim ({token_dim}) != dim ({dim}) when skip_token_embedding is True"
+                )
+
+        self.pos_embedding = nn.Parameter(torch.randn(1, num_tokens, dim))
+        if emb_dropout_type == "drop":
+            self.dropout = DropTokenDropout(emb_dropout)
+        elif emb_dropout_type == "zero":
+            self.dropout = ZeroTokenDropout(emb_dropout)
+        elif emb_dropout_type == "normal":
+            self.dropout = nn.Dropout(emb_dropout)
+
+        self.transformer = TransformerCrossAttn(
+            dim,
+            depth,
+            heads,
+            dim_head,
+            mlp_dim,
+            dropout,
+            norm=norm,
+            norm_cond_dim=norm_cond_dim,
+            context_dim=context_dim,
+        )
+
+    def forward(self, inp: torch.Tensor, *args, context=None, context_list=None):
+        x = self.to_token_embedding(inp)
+        b, n, _ = x.shape
+
+        x = self.dropout(x)
+        x += self.pos_embedding[:, :n]
+
+        x = self.transformer(x, *args, context=context, context_list=context_list)
+        return x
+

hmr2/models/components/t_cond_mlp.py

@@ -0,0 +1,199 @@
+import copy
+from typing import List, Optional
+
+import torch
+
+
+class AdaptiveLayerNorm1D(torch.nn.Module):
+    def __init__(self, data_dim: int, norm_cond_dim: int):
+        super().__init__()
+        if data_dim <= 0:
+            raise ValueError(f"data_dim must be positive, but got {data_dim}")
+        if norm_cond_dim <= 0:
+            raise ValueError(f"norm_cond_dim must be positive, but got {norm_cond_dim}")
+        self.norm = torch.nn.LayerNorm(
+            data_dim
+        )  # TODO: Check if elementwise_affine=True is correct
+        self.linear = torch.nn.Linear(norm_cond_dim, 2 * data_dim)
+        torch.nn.init.zeros_(self.linear.weight)
+        torch.nn.init.zeros_(self.linear.bias)
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
+        # x: (batch, ..., data_dim)
+        # t: (batch, norm_cond_dim)
+        # return: (batch, data_dim)
+        x = self.norm(x)
+        alpha, beta = self.linear(t).chunk(2, dim=-1)
+
+        # Add singleton dimensions to alpha and beta
+        if x.dim() > 2:
+            alpha = alpha.view(alpha.shape[0], *([1] * (x.dim() - 2)), alpha.shape[1])
+            beta = beta.view(beta.shape[0], *([1] * (x.dim() - 2)), beta.shape[1])
+
+        return x * (1 + alpha) + beta
+
+
+class SequentialCond(torch.nn.Sequential):
+    def forward(self, input, *args, **kwargs):
+        for module in self:
+            if isinstance(module, (AdaptiveLayerNorm1D, SequentialCond, ResidualMLPBlock)):
+                # print(f'Passing on args to {module}', [a.shape for a in args])
+                input = module(input, *args, **kwargs)
+            else:
+                # print(f'Skipping passing args to {module}', [a.shape for a in args])
+                input = module(input)
+        return input
+
+
+def normalization_layer(norm: Optional[str], dim: int, norm_cond_dim: int = -1):
+    if norm == "batch":
+        return torch.nn.BatchNorm1d(dim)
+    elif norm == "layer":
+        return torch.nn.LayerNorm(dim)
+    elif norm == "ada":
+        assert norm_cond_dim > 0, f"norm_cond_dim must be positive, got {norm_cond_dim}"
+        return AdaptiveLayerNorm1D(dim, norm_cond_dim)
+    elif norm is None:
+        return torch.nn.Identity()
+    else:
+        raise ValueError(f"Unknown norm: {norm}")
+
+
+def linear_norm_activ_dropout(
+    input_dim: int,
+    output_dim: int,
+    activation: torch.nn.Module = torch.nn.ReLU(),
+    bias: bool = True,
+    norm: Optional[str] = "layer",  # Options: ada/batch/layer
+    dropout: float = 0.0,
+    norm_cond_dim: int = -1,
+) -> SequentialCond:
+    layers = []
+    layers.append(torch.nn.Linear(input_dim, output_dim, bias=bias))
+    if norm is not None:
+        layers.append(normalization_layer(norm, output_dim, norm_cond_dim))
+    layers.append(copy.deepcopy(activation))
+    if dropout > 0.0:
+        layers.append(torch.nn.Dropout(dropout))
+    return SequentialCond(*layers)
+
+
+def create_simple_mlp(
+    input_dim: int,
+    hidden_dims: List[int],
+    output_dim: int,
+    activation: torch.nn.Module = torch.nn.ReLU(),
+    bias: bool = True,
+    norm: Optional[str] = "layer",  # Options: ada/batch/layer
+    dropout: float = 0.0,
+    norm_cond_dim: int = -1,
+) -> SequentialCond:
+    layers = []
+    prev_dim = input_dim
+    for hidden_dim in hidden_dims:
+        layers.extend(
+            linear_norm_activ_dropout(
+                prev_dim, hidden_dim, activation, bias, norm, dropout, norm_cond_dim
+            )
+        )
+        prev_dim = hidden_dim
+    layers.append(torch.nn.Linear(prev_dim, output_dim, bias=bias))
+    return SequentialCond(*layers)
+
+
+class ResidualMLPBlock(torch.nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        num_hidden_layers: int,
+        output_dim: int,
+        activation: torch.nn.Module = torch.nn.ReLU(),
+        bias: bool = True,
+        norm: Optional[str] = "layer",  # Options: ada/batch/layer
+        dropout: float = 0.0,
+        norm_cond_dim: int = -1,
+    ):
+        super().__init__()
+        if not (input_dim == output_dim == hidden_dim):
+            raise NotImplementedError(
+                f"input_dim {input_dim} != output_dim {output_dim} is not implemented"
+            )
+
+        layers = []
+        prev_dim = input_dim
+        for i in range(num_hidden_layers):
+            layers.append(
+                linear_norm_activ_dropout(
+                    prev_dim, hidden_dim, activation, bias, norm, dropout, norm_cond_dim
+                )
+            )
+            prev_dim = hidden_dim
+        self.model = SequentialCond(*layers)
+        self.skip = torch.nn.Identity()
+
+    def forward(self, x: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        return x + self.model(x, *args, **kwargs)
+
+
+class ResidualMLP(torch.nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        num_hidden_layers: int,
+        output_dim: int,
+        activation: torch.nn.Module = torch.nn.ReLU(),
+        bias: bool = True,
+        norm: Optional[str] = "layer",  # Options: ada/batch/layer
+        dropout: float = 0.0,
+        num_blocks: int = 1,
+        norm_cond_dim: int = -1,
+    ):
+        super().__init__()
+        self.input_dim = input_dim
+        self.model = SequentialCond(
+            linear_norm_activ_dropout(
+                input_dim, hidden_dim, activation, bias, norm, dropout, norm_cond_dim
+            ),
+            *[
+                ResidualMLPBlock(
+                    hidden_dim,
+                    hidden_dim,
+                    num_hidden_layers,
+                    hidden_dim,
+                    activation,
+                    bias,
+                    norm,
+                    dropout,
+                    norm_cond_dim,
+                )
+                for _ in range(num_blocks)
+            ],
+            torch.nn.Linear(hidden_dim, output_dim, bias=bias),
+        )
+
+    def forward(self, x: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        return self.model(x, *args, **kwargs)
+
+
+class FrequencyEmbedder(torch.nn.Module):
+    def __init__(self, num_frequencies, max_freq_log2):
+        super().__init__()
+        frequencies = 2 ** torch.linspace(0, max_freq_log2, steps=num_frequencies)
+        self.register_buffer("frequencies", frequencies)
+
+    def forward(self, x):
+        # x should be of size (N,) or (N, D)
+        N = x.size(0)
+        if x.dim() == 1:  # (N,)
+            x = x.unsqueeze(1)  # (N, D) where D=1
+        x_unsqueezed = x.unsqueeze(-1)  # (N, D, 1)
+        scaled = self.frequencies.view(1, 1, -1) * x_unsqueezed  # (N, D, num_frequencies)
+        s = torch.sin(scaled)
+        c = torch.cos(scaled)
+        embedded = torch.cat([s, c, x_unsqueezed], dim=-1).view(
+            N, -1
+        )  # (N, D * 2 * num_frequencies + D)
+        return embedded
+

hmr2/models/discriminator.py

@@ -0,0 +1,99 @@
+import torch
+import torch.nn as nn
+
+class Discriminator(nn.Module):
+
+    def __init__(self):
+        """
+        Pose + Shape discriminator proposed in HMR
+        """
+        super(Discriminator, self).__init__()
+
+        self.num_joints = 23
+        # poses_alone
+        self.D_conv1 = nn.Conv2d(9, 32, kernel_size=1)
+        nn.init.xavier_uniform_(self.D_conv1.weight)
+        nn.init.zeros_(self.D_conv1.bias)
+        self.relu = nn.ReLU(inplace=True)
+        self.D_conv2 = nn.Conv2d(32, 32, kernel_size=1)
+        nn.init.xavier_uniform_(self.D_conv2.weight)
+        nn.init.zeros_(self.D_conv2.bias)
+        pose_out = []
+        for i in range(self.num_joints):
+            pose_out_temp = nn.Linear(32, 1)
+            nn.init.xavier_uniform_(pose_out_temp.weight)
+            nn.init.zeros_(pose_out_temp.bias)
+            pose_out.append(pose_out_temp)
+        self.pose_out = nn.ModuleList(pose_out)
+
+        # betas
+        self.betas_fc1 = nn.Linear(10, 10)
+        nn.init.xavier_uniform_(self.betas_fc1.weight)
+        nn.init.zeros_(self.betas_fc1.bias)
+        self.betas_fc2 = nn.Linear(10, 5)
+        nn.init.xavier_uniform_(self.betas_fc2.weight)
+        nn.init.zeros_(self.betas_fc2.bias)
+        self.betas_out = nn.Linear(5, 1)
+        nn.init.xavier_uniform_(self.betas_out.weight)
+        nn.init.zeros_(self.betas_out.bias)
+
+        # poses_joint
+        self.D_alljoints_fc1 = nn.Linear(32*self.num_joints, 1024)
+        nn.init.xavier_uniform_(self.D_alljoints_fc1.weight)
+        nn.init.zeros_(self.D_alljoints_fc1.bias)
+        self.D_alljoints_fc2 = nn.Linear(1024, 1024)
+        nn.init.xavier_uniform_(self.D_alljoints_fc2.weight)
+        nn.init.zeros_(self.D_alljoints_fc2.bias)
+        self.D_alljoints_out = nn.Linear(1024, 1)
+        nn.init.xavier_uniform_(self.D_alljoints_out.weight)
+        nn.init.zeros_(self.D_alljoints_out.bias)
+
+
+    def forward(self, poses: torch.Tensor, betas: torch.Tensor) -> torch.Tensor:
+        """
+        Forward pass of the discriminator.
+        Args:
+            poses (torch.Tensor): Tensor of shape (B, 23, 3, 3) containing a batch of SMPL body poses (excluding the global orientation).
+            betas (torch.Tensor): Tensor of shape (B, 10) containign a batch of SMPL beta coefficients.
+        Returns:
+            torch.Tensor: Discriminator output with shape (B, 25)
+        """
+        #import ipdb; ipdb.set_trace()
+        #bn = poses.shape[0]
+        # poses B x 207
+        #poses = poses.reshape(bn, -1)
+        # poses B x num_joints x 1 x 9
+        poses = poses.reshape(-1, self.num_joints, 1, 9)
+        bn = poses.shape[0]
+        # poses B x 9 x num_joints x 1
+        poses = poses.permute(0, 3, 1, 2).contiguous()
+
+        # poses_alone
+        poses = self.D_conv1(poses)
+        poses = self.relu(poses)
+        poses = self.D_conv2(poses)
+        poses = self.relu(poses)
+
+        poses_out = []
+        for i in range(self.num_joints):
+            poses_out_ = self.pose_out[i](poses[:, :, i, 0])
+            poses_out.append(poses_out_)
+        poses_out = torch.cat(poses_out, dim=1)
+
+        # betas
+        betas = self.betas_fc1(betas)
+        betas = self.relu(betas)
+        betas = self.betas_fc2(betas)
+        betas = self.relu(betas)
+        betas_out = self.betas_out(betas)
+
+        # poses_joint
+        poses = poses.reshape(bn,-1)
+        poses_all = self.D_alljoints_fc1(poses)
+        poses_all = self.relu(poses_all)
+        poses_all = self.D_alljoints_fc2(poses_all)
+        poses_all = self.relu(poses_all)
+        poses_all_out = self.D_alljoints_out(poses_all)
+
+        disc_out = torch.cat((poses_out, betas_out, poses_all_out), 1)
+        return disc_out

hmr2/models/heads/__init__.py

@@ -0,0 +1 @@
+from .smpl_head import build_smpl_head

hmr2/models/heads/smpl_head.py

@@ -0,0 +1,111 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+import einops
+
+from ...utils.geometry import rot6d_to_rotmat, aa_to_rotmat
+from ..components.pose_transformer import TransformerDecoder
+
+def build_smpl_head(cfg):
+    smpl_head_type = cfg.MODEL.SMPL_HEAD.get('TYPE', 'hmr')
+    if  smpl_head_type == 'transformer_decoder':
+        return SMPLTransformerDecoderHead(cfg)
+    else:
+        raise ValueError('Unknown SMPL head type: {}'.format(smpl_head_type))
+
+class SMPLTransformerDecoderHead(nn.Module):
+    """ Cross-attention based SMPL Transformer decoder
+    """
+
+    def __init__(self, cfg):
+        super().__init__()
+        self.cfg = cfg
+        self.joint_rep_type = cfg.MODEL.SMPL_HEAD.get('JOINT_REP', '6d')
+        self.joint_rep_dim = {'6d': 6, 'aa': 3}[self.joint_rep_type]
+        npose = self.joint_rep_dim * (cfg.SMPL.NUM_BODY_JOINTS + 1)
+        self.npose = npose
+        self.input_is_mean_shape = cfg.MODEL.SMPL_HEAD.get('TRANSFORMER_INPUT', 'zero') == 'mean_shape'
+        transformer_args = dict(
+            num_tokens=1,
+            token_dim=(npose + 10 + 3) if self.input_is_mean_shape else 1,
+            dim=1024,
+        )
+        transformer_args = (transformer_args | dict(cfg.MODEL.SMPL_HEAD.TRANSFORMER_DECODER))
+        self.transformer = TransformerDecoder(
+            **transformer_args
+        )
+        dim=transformer_args['dim']
+        self.decpose = nn.Linear(dim, npose)
+        self.decshape = nn.Linear(dim, 10)
+        self.deccam = nn.Linear(dim, 3)
+
+        if cfg.MODEL.SMPL_HEAD.get('INIT_DECODER_XAVIER', False):
+            # True by default in MLP. False by default in Transformer
+            nn.init.xavier_uniform_(self.decpose.weight, gain=0.01)
+            nn.init.xavier_uniform_(self.decshape.weight, gain=0.01)
+            nn.init.xavier_uniform_(self.deccam.weight, gain=0.01)
+
+        mean_params = np.load(cfg.SMPL.MEAN_PARAMS)
+        init_body_pose = torch.from_numpy(mean_params['pose'].astype(np.float32)).unsqueeze(0)
+        init_betas = torch.from_numpy(mean_params['shape'].astype('float32')).unsqueeze(0)
+        init_cam = torch.from_numpy(mean_params['cam'].astype(np.float32)).unsqueeze(0)
+        self.register_buffer('init_body_pose', init_body_pose)
+        self.register_buffer('init_betas', init_betas)
+        self.register_buffer('init_cam', init_cam)
+
+    def forward(self, x, **kwargs):
+
+        batch_size = x.shape[0]
+        # vit pretrained backbone is channel-first. Change to token-first
+        x = einops.rearrange(x, 'b c h w -> b (h w) c')
+
+        init_body_pose = self.init_body_pose.expand(batch_size, -1)
+        init_betas = self.init_betas.expand(batch_size, -1)
+        init_cam = self.init_cam.expand(batch_size, -1)
+
+        # TODO: Convert init_body_pose to aa rep if needed
+        if self.joint_rep_type == 'aa':
+            raise NotImplementedError
+
+        pred_body_pose = init_body_pose
+        pred_betas = init_betas
+        pred_cam = init_cam
+        pred_body_pose_list = []
+        pred_betas_list = []
+        pred_cam_list = []
+        for i in range(self.cfg.MODEL.SMPL_HEAD.get('IEF_ITERS', 1)):
+            # Input token to transformer is zero token
+            if self.input_is_mean_shape:
+                token = torch.cat([pred_body_pose, pred_betas, pred_cam], dim=1)[:,None,:]
+            else:
+                token = torch.zeros(batch_size, 1, 1).to(x.device)
+
+            # Pass through transformer
+            token_out = self.transformer(token, context=x)
+            token_out = token_out.squeeze(1) # (B, C)
+
+            # Readout from token_out
+            pred_body_pose = self.decpose(token_out) + pred_body_pose
+            pred_betas = self.decshape(token_out) + pred_betas
+            pred_cam = self.deccam(token_out) + pred_cam
+            pred_body_pose_list.append(pred_body_pose)
+            pred_betas_list.append(pred_betas)
+            pred_cam_list.append(pred_cam)
+
+        # Convert self.joint_rep_type -> rotmat
+        joint_conversion_fn = {
+            '6d': rot6d_to_rotmat,
+            'aa': lambda x: aa_to_rotmat(x.view(-1, 3).contiguous())
+        }[self.joint_rep_type]
+
+        pred_smpl_params_list = {}
+        pred_smpl_params_list['body_pose'] = torch.cat([joint_conversion_fn(pbp).view(batch_size, -1, 3, 3)[:, 1:, :, :] for pbp in pred_body_pose_list], dim=0)
+        pred_smpl_params_list['betas'] = torch.cat(pred_betas_list, dim=0)
+        pred_smpl_params_list['cam'] = torch.cat(pred_cam_list, dim=0)
+        pred_body_pose = joint_conversion_fn(pred_body_pose).view(batch_size, self.cfg.SMPL.NUM_BODY_JOINTS+1, 3, 3)
+
+        pred_smpl_params = {'global_orient': pred_body_pose[:, [0]],
+                            'body_pose': pred_body_pose[:, 1:],
+                            'betas': pred_betas}
+        return pred_smpl_params, pred_cam, pred_smpl_params_list

hmr2/models/hmr2.py

@@ -0,0 +1,363 @@
+import torch
+import pytorch_lightning as pl
+from typing import Any, Dict, Mapping, Tuple
+
+from yacs.config import CfgNode
+
+from ..utils import SkeletonRenderer, MeshRenderer
+from ..utils.geometry import aa_to_rotmat, perspective_projection
+from .backbones import create_backbone
+from .heads import build_smpl_head
+from .discriminator import Discriminator
+from .losses import Keypoint3DLoss, Keypoint2DLoss, ParameterLoss
+from . import SMPL
+
+
+class HMR2(pl.LightningModule):
+
+    def __init__(self, cfg: CfgNode, init_renderer: bool = True):
+        """
+        Setup HMR2 model
+        Args:
+            cfg (CfgNode): Config file as a yacs CfgNode
+        """
+        super().__init__()
+
+        # Save hyperparameters
+        self.save_hyperparameters(logger=False, ignore=['init_renderer'])
+
+        self.cfg = cfg
+        # Create backbone feature extractor
+        self.backbone = create_backbone(cfg)
+        
+        # Create SMPL head
+        self.smpl_head = build_smpl_head(cfg)
+
+        # Create discriminator
+        if self.cfg.LOSS_WEIGHTS.ADVERSARIAL > 0:
+            self.discriminator = Discriminator()
+
+        # Define loss functions
+        self.keypoint_3d_loss = Keypoint3DLoss(loss_type='l1')
+        self.keypoint_2d_loss = Keypoint2DLoss(loss_type='l1')
+        self.smpl_parameter_loss = ParameterLoss()
+
+        # Instantiate SMPL model
+        smpl_cfg = {k.lower(): v for k,v in dict(cfg.SMPL).items()}
+        self.smpl = SMPL(**smpl_cfg)
+
+        # Buffer that shows whetheer we need to initialize ActNorm layers
+        self.register_buffer('initialized', torch.tensor(False))
+        # Setup renderer for visualization
+        if init_renderer:
+            self.renderer = SkeletonRenderer(self.cfg)
+            self.mesh_renderer = MeshRenderer(self.cfg, faces=self.smpl.faces)
+        else:
+            self.renderer = None
+            self.mesh_renderer = None
+
+        # Disable automatic optimization since we use adversarial training
+        self.automatic_optimization = False
+
+    def get_parameters(self):
+        all_params = list(self.smpl_head.parameters())
+        all_params += list(self.backbone.parameters())
+        return all_params
+
+    def configure_optimizers(self) -> Tuple[torch.optim.Optimizer, torch.optim.Optimizer]:
+        """
+        Setup model and distriminator Optimizers
+        Returns:
+            Tuple[torch.optim.Optimizer, torch.optim.Optimizer]: Model and discriminator optimizers
+        """
+        param_groups = [{'params': filter(lambda p: p.requires_grad, self.get_parameters()), 'lr': self.cfg.TRAIN.LR}]
+
+        optimizer = torch.optim.AdamW(params=param_groups,
+                                        # lr=self.cfg.TRAIN.LR,
+                                        weight_decay=self.cfg.TRAIN.WEIGHT_DECAY)
+        optimizer_disc = torch.optim.AdamW(params=self.discriminator.parameters(),
+                                            lr=self.cfg.TRAIN.LR,
+                                            weight_decay=self.cfg.TRAIN.WEIGHT_DECAY)
+
+        return optimizer, optimizer_disc
+
+    def forward_step(self, batch: Dict, train: bool = False) -> Dict:
+        """
+        Run a forward step of the network
+        Args:
+            batch (Dict): Dictionary containing batch data
+            train (bool): Flag indicating whether it is training or validation mode
+        Returns:
+            Dict: Dictionary containing the regression output
+        """
+
+        # Use RGB image as input
+        x = batch['img']
+        batch_size = x.shape[0]
+
+        # Compute conditioning features using the backbone
+        # if using ViT backbone, we need to use a different aspect ratio
+        conditioning_feats = self.backbone(x[:,:,:,32:-32])
+
+        pred_smpl_params, pred_cam, _ = self.smpl_head(conditioning_feats)
+
+        # Store useful regression outputs to the output dict
+        output = {}
+        output['pred_cam'] = pred_cam
+        output['pred_smpl_params'] = {k: v.clone() for k,v in pred_smpl_params.items()}
+
+        # Compute camera translation
+        device = pred_smpl_params['body_pose'].device
+        dtype = pred_smpl_params['body_pose'].dtype
+        focal_length = self.cfg.EXTRA.FOCAL_LENGTH * torch.ones(batch_size, 2, device=device, dtype=dtype)
+        pred_cam_t = torch.stack([pred_cam[:, 1],
+                                  pred_cam[:, 2],
+                                  2*focal_length[:, 0]/(self.cfg.MODEL.IMAGE_SIZE * pred_cam[:, 0] +1e-9)],dim=-1)
+        output['pred_cam_t'] = pred_cam_t
+        output['focal_length'] = focal_length
+
+        # Compute model vertices, joints and the projected joints
+        pred_smpl_params['global_orient'] = pred_smpl_params['global_orient'].reshape(batch_size, -1, 3, 3)
+        pred_smpl_params['body_pose'] = pred_smpl_params['body_pose'].reshape(batch_size, -1, 3, 3)
+        pred_smpl_params['betas'] = pred_smpl_params['betas'].reshape(batch_size, -1)
+        smpl_output = self.smpl(**{k: v.float() for k,v in pred_smpl_params.items()}, pose2rot=False)
+        pred_keypoints_3d = smpl_output.joints
+        pred_vertices = smpl_output.vertices
+        output['pred_keypoints_3d'] = pred_keypoints_3d.reshape(batch_size, -1, 3)
+        output['pred_vertices'] = pred_vertices.reshape(batch_size, -1, 3)
+        pred_cam_t = pred_cam_t.reshape(-1, 3)
+        focal_length = focal_length.reshape(-1, 2)
+        pred_keypoints_2d = perspective_projection(pred_keypoints_3d,
+                                                   translation=pred_cam_t,
+                                                   focal_length=focal_length / self.cfg.MODEL.IMAGE_SIZE)
+
+        output['pred_keypoints_2d'] = pred_keypoints_2d.reshape(batch_size, -1, 2)
+        return output
+
+    def compute_loss(self, batch: Dict, output: Dict, train: bool = True) -> torch.Tensor:
+        """
+        Compute losses given the input batch and the regression output
+        Args:
+            batch (Dict): Dictionary containing batch data
+            output (Dict): Dictionary containing the regression output
+            train (bool): Flag indicating whether it is training or validation mode
+        Returns:
+            torch.Tensor : Total loss for current batch
+        """
+
+        pred_smpl_params = output['pred_smpl_params']
+        pred_keypoints_2d = output['pred_keypoints_2d']
+        pred_keypoints_3d = output['pred_keypoints_3d']
+
+
+        batch_size = pred_smpl_params['body_pose'].shape[0]
+        device = pred_smpl_params['body_pose'].device
+        dtype = pred_smpl_params['body_pose'].dtype
+
+        # Get annotations
+        gt_keypoints_2d = batch['keypoints_2d']
+        gt_keypoints_3d = batch['keypoints_3d']
+        gt_smpl_params = batch['smpl_params']
+        has_smpl_params = batch['has_smpl_params']
+        is_axis_angle = batch['smpl_params_is_axis_angle']
+
+        # Compute 3D keypoint loss
+        loss_keypoints_2d = self.keypoint_2d_loss(pred_keypoints_2d, gt_keypoints_2d)
+        loss_keypoints_3d = self.keypoint_3d_loss(pred_keypoints_3d, gt_keypoints_3d, pelvis_id=25+14)
+
+        # Compute loss on SMPL parameters
+        loss_smpl_params = {}
+        for k, pred in pred_smpl_params.items():
+            gt = gt_smpl_params[k].view(batch_size, -1)
+            if is_axis_angle[k].all():
+                gt = aa_to_rotmat(gt.reshape(-1, 3)).view(batch_size, -1, 3, 3)
+            has_gt = has_smpl_params[k]
+            loss_smpl_params[k] = self.smpl_parameter_loss(pred.reshape(batch_size, -1), gt.reshape(batch_size, -1), has_gt)
+
+        # # Filter out images with corresponding SMPL parameter annotations
+        # smpl_params = {k: v.clone() for k,v in gt_smpl_params.items()}
+        # smpl_params['body_pose'] = aa_to_rotmat(smpl_params['body_pose'].reshape(-1, 3)).reshape(batch_size, -1, 3, 3)[:, :, :, :2].permute(0, 1, 3, 2).reshape(batch_size, -1)
+        # smpl_params['global_orient'] = aa_to_rotmat(smpl_params['global_orient'].reshape(-1, 3)).reshape(batch_size, -1, 3, 3)[:, :, :, :2].permute(0, 1, 3, 2).reshape(batch_size, -1)
+        # smpl_params['betas'] = smpl_params['betas']
+        # has_smpl_params = (batch['has_smpl_params']['body_pose'] > 0)
+        # smpl_params = {k: v[has_smpl_params] for k, v in smpl_params.items()}
+
+        loss = self.cfg.LOSS_WEIGHTS['KEYPOINTS_3D'] * loss_keypoints_3d+\
+               self.cfg.LOSS_WEIGHTS['KEYPOINTS_2D'] * loss_keypoints_2d+\
+               sum([loss_smpl_params[k] * self.cfg.LOSS_WEIGHTS[k.upper()] for k in loss_smpl_params])
+
+        losses = dict(loss=loss.detach(),
+                      loss_keypoints_2d=loss_keypoints_2d.detach(),
+                      loss_keypoints_3d=loss_keypoints_3d.detach())
+
+        for k, v in loss_smpl_params.items():
+            losses['loss_' + k] = v.detach()
+
+        output['losses'] = losses
+
+        return loss
+
+    # Tensoroboard logging should run from first rank only
+    @pl.utilities.rank_zero.rank_zero_only
+    def tensorboard_logging(self, batch: Dict, output: Dict, step_count: int, train: bool = True, write_to_summary_writer: bool = True) -> None:
+        """
+        Log results to Tensorboard
+        Args:
+            batch (Dict): Dictionary containing batch data
+            output (Dict): Dictionary containing the regression output
+            step_count (int): Global training step count
+            train (bool): Flag indicating whether it is training or validation mode
+        """
+
+        mode = 'train' if train else 'val'
+        batch_size = batch['keypoints_2d'].shape[0]
+        images = batch['img']
+        images = images * torch.tensor([0.229, 0.224, 0.225], device=images.device).reshape(1,3,1,1)
+        images = images + torch.tensor([0.485, 0.456, 0.406], device=images.device).reshape(1,3,1,1)
+        #images = 255*images.permute(0, 2, 3, 1).cpu().numpy()
+
+        pred_keypoints_3d = output['pred_keypoints_3d'].detach().reshape(batch_size, -1, 3)
+        pred_vertices = output['pred_vertices'].detach().reshape(batch_size, -1, 3)
+        focal_length = output['focal_length'].detach().reshape(batch_size, 2)
+        gt_keypoints_3d = batch['keypoints_3d']
+        gt_keypoints_2d = batch['keypoints_2d']
+        losses = output['losses']
+        pred_cam_t = output['pred_cam_t'].detach().reshape(batch_size, 3)
+        pred_keypoints_2d = output['pred_keypoints_2d'].detach().reshape(batch_size, -1, 2)
+
+        if write_to_summary_writer:
+            summary_writer = self.logger.experiment
+            for loss_name, val in losses.items():
+                summary_writer.add_scalar(mode +'/' + loss_name, val.detach().item(), step_count)
+        num_images = min(batch_size, self.cfg.EXTRA.NUM_LOG_IMAGES)
+
+        gt_keypoints_3d = batch['keypoints_3d']
+        pred_keypoints_3d = output['pred_keypoints_3d'].detach().reshape(batch_size, -1, 3)
+
+        # We render the skeletons instead of the full mesh because rendering a lot of meshes will make the training slow.
+        #predictions = self.renderer(pred_keypoints_3d[:num_images],
+        #                            gt_keypoints_3d[:num_images],
+        #                            2 * gt_keypoints_2d[:num_images],
+        #                            images=images[:num_images],
+        #                            camera_translation=pred_cam_t[:num_images])
+        predictions = self.mesh_renderer.visualize_tensorboard(pred_vertices[:num_images].cpu().numpy(),
+                                                               pred_cam_t[:num_images].cpu().numpy(),
+                                                               images[:num_images].cpu().numpy(),
+                                                               pred_keypoints_2d[:num_images].cpu().numpy(),
+                                                               gt_keypoints_2d[:num_images].cpu().numpy(),
+                                                               focal_length=focal_length[:num_images].cpu().numpy())
+        if write_to_summary_writer:
+            summary_writer.add_image('%s/predictions' % mode, predictions, step_count)
+
+        return predictions
+
+    def forward(self, batch: Dict) -> Dict:
+        """
+        Run a forward step of the network in val mode
+        Args:
+            batch (Dict): Dictionary containing batch data
+        Returns:
+            Dict: Dictionary containing the regression output
+        """
+        return self.forward_step(batch, train=False)
+
+    def training_step_discriminator(self, batch: Dict,
+                                    body_pose: torch.Tensor,
+                                    betas: torch.Tensor,
+                                    optimizer: torch.optim.Optimizer) -> torch.Tensor:
+        """
+        Run a discriminator training step
+        Args:
+            batch (Dict): Dictionary containing mocap batch data
+            body_pose (torch.Tensor): Regressed body pose from current step
+            betas (torch.Tensor): Regressed betas from current step
+            optimizer (torch.optim.Optimizer): Discriminator optimizer
+        Returns:
+            torch.Tensor: Discriminator loss
+        """
+        batch_size = body_pose.shape[0]
+        gt_body_pose = batch['body_pose']
+        gt_betas = batch['betas']
+        gt_rotmat = aa_to_rotmat(gt_body_pose.view(-1,3)).view(batch_size, -1, 3, 3)
+        disc_fake_out = self.discriminator(body_pose.detach(), betas.detach())
+        loss_fake = ((disc_fake_out - 0.0) ** 2).sum() / batch_size
+        disc_real_out = self.discriminator(gt_rotmat, gt_betas)
+        loss_real = ((disc_real_out - 1.0) ** 2).sum() / batch_size
+        loss_disc = loss_fake + loss_real
+        loss = self.cfg.LOSS_WEIGHTS.ADVERSARIAL * loss_disc
+        optimizer.zero_grad()
+        self.manual_backward(loss)
+        optimizer.step()
+        return loss_disc.detach()
+
+    def training_step(self, joint_batch: Dict, batch_idx: int) -> Dict:
+        """
+        Run a full training step
+        Args:
+            joint_batch (Dict): Dictionary containing image and mocap batch data
+            batch_idx (int): Unused.
+            batch_idx (torch.Tensor): Unused.
+        Returns:
+            Dict: Dictionary containing regression output.
+        """
+        batch = joint_batch['img']
+        mocap_batch = joint_batch['mocap']
+        optimizer = self.optimizers(use_pl_optimizer=True)
+        if self.cfg.LOSS_WEIGHTS.ADVERSARIAL > 0:
+            optimizer, optimizer_disc = optimizer
+
+        # Update learning rates
+        self.update_learning_rates(batch_idx)
+
+        batch_size = batch['img'].shape[0]
+        output = self.forward_step(batch, train=True)
+        pred_smpl_params = output['pred_smpl_params']
+        if self.cfg.get('UPDATE_GT_SPIN', False):
+            self.update_batch_gt_spin(batch, output)
+        loss = self.compute_loss(batch, output, train=True)
+        if self.cfg.LOSS_WEIGHTS.ADVERSARIAL > 0:
+            disc_out = self.discriminator(pred_smpl_params['body_pose'].reshape(batch_size, -1), pred_smpl_params['betas'].reshape(batch_size, -1))
+            loss_adv = ((disc_out - 1.0) ** 2).sum() / batch_size
+            loss = loss + self.cfg.LOSS_WEIGHTS.ADVERSARIAL * loss_adv
+
+        # Error if Nan
+        if torch.isnan(loss):
+            raise ValueError('Loss is NaN')
+
+        optimizer.zero_grad()
+        self.manual_backward(loss)
+        # Clip gradient
+        if self.cfg.TRAIN.get('GRAD_CLIP_VAL', 0) > 0:
+            gn = torch.nn.utils.clip_grad_norm_(self.get_parameters(), self.cfg.TRAIN.GRAD_CLIP_VAL, error_if_nonfinite=True)
+            self.log('train/grad_norm', gn, on_step=True, on_epoch=True, prog_bar=True, logger=True)
+        optimizer.step()
+        if self.cfg.LOSS_WEIGHTS.ADVERSARIAL > 0:
+            loss_disc = self.training_step_discriminator(mocap_batch, pred_smpl_params['body_pose'].reshape(batch_size, -1), pred_smpl_params['betas'].reshape(batch_size, -1), optimizer_disc)
+            output['losses']['loss_gen'] = loss_adv
+            output['losses']['loss_disc'] = loss_disc
+
+        if self.global_step > 0 and self.global_step % self.cfg.GENERAL.LOG_STEPS == 0:
+            self.tensorboard_logging(batch, output, self.global_step, train=True)
+
+        self.log('train/loss', output['losses']['loss'], on_step=True, on_epoch=True, prog_bar=True, logger=False)
+
+        return output
+
+    def validation_step(self, batch: Dict, batch_idx: int, dataloader_idx=0) -> Dict:
+        """
+        Run a validation step and log to Tensorboard
+        Args:
+            batch (Dict): Dictionary containing batch data
+            batch_idx (int): Unused.
+        Returns:
+            Dict: Dictionary containing regression output.
+        """
+        # batch_size = batch['img'].shape[0]
+        output = self.forward_step(batch, train=False)
+
+        pred_smpl_params = output['pred_smpl_params']
+        loss = self.compute_loss(batch, output, train=False)
+        output['loss'] = loss
+        self.tensorboard_logging(batch, output, self.global_step, train=False)
+
+        return output

hmr2/models/losses.py

@@ -0,0 +1,92 @@
+import torch
+import torch.nn as nn
+
+class Keypoint2DLoss(nn.Module):
+
+    def __init__(self, loss_type: str = 'l1'):
+        """
+        2D keypoint loss module.
+        Args:
+            loss_type (str): Choose between l1 and l2 losses.
+        """
+        super(Keypoint2DLoss, self).__init__()
+        if loss_type == 'l1':
+            self.loss_fn = nn.L1Loss(reduction='none')
+        elif loss_type == 'l2':
+            self.loss_fn = nn.MSELoss(reduction='none')
+        else:
+            raise NotImplementedError('Unsupported loss function')
+
+    def forward(self, pred_keypoints_2d: torch.Tensor, gt_keypoints_2d: torch.Tensor) -> torch.Tensor:
+        """
+        Compute 2D reprojection loss on the keypoints.
+        Args:
+            pred_keypoints_2d (torch.Tensor): Tensor of shape [B, S, N, 2] containing projected 2D keypoints (B: batch_size, S: num_samples, N: num_keypoints)
+            gt_keypoints_2d (torch.Tensor): Tensor of shape [B, S, N, 3] containing the ground truth 2D keypoints and confidence.
+        Returns:
+            torch.Tensor: 2D keypoint loss.
+        """
+        conf = gt_keypoints_2d[:, :, -1].unsqueeze(-1).clone()
+        batch_size = conf.shape[0]
+        loss = (conf * self.loss_fn(pred_keypoints_2d, gt_keypoints_2d[:, :, :-1])).sum(dim=(1,2))
+        return loss.sum()
+
+
+class Keypoint3DLoss(nn.Module):
+
+    def __init__(self, loss_type: str = 'l1'):
+        """
+        3D keypoint loss module.
+        Args:
+            loss_type (str): Choose between l1 and l2 losses.
+        """
+        super(Keypoint3DLoss, self).__init__()
+        if loss_type == 'l1':
+            self.loss_fn = nn.L1Loss(reduction='none')
+        elif loss_type == 'l2':
+            self.loss_fn = nn.MSELoss(reduction='none')
+        else:
+            raise NotImplementedError('Unsupported loss function')
+
+    def forward(self, pred_keypoints_3d: torch.Tensor, gt_keypoints_3d: torch.Tensor, pelvis_id: int = 39):
+        """
+        Compute 3D keypoint loss.
+        Args:
+            pred_keypoints_3d (torch.Tensor): Tensor of shape [B, S, N, 3] containing the predicted 3D keypoints (B: batch_size, S: num_samples, N: num_keypoints)
+            gt_keypoints_3d (torch.Tensor): Tensor of shape [B, S, N, 4] containing the ground truth 3D keypoints and confidence.
+        Returns:
+            torch.Tensor: 3D keypoint loss.
+        """
+        batch_size = pred_keypoints_3d.shape[0]
+        gt_keypoints_3d = gt_keypoints_3d.clone()
+        pred_keypoints_3d = pred_keypoints_3d - pred_keypoints_3d[:, pelvis_id, :].unsqueeze(dim=1)
+        gt_keypoints_3d[:, :, :-1] = gt_keypoints_3d[:, :, :-1] - gt_keypoints_3d[:, pelvis_id, :-1].unsqueeze(dim=1)
+        conf = gt_keypoints_3d[:, :, -1].unsqueeze(-1).clone()
+        gt_keypoints_3d = gt_keypoints_3d[:, :, :-1]
+        loss = (conf * self.loss_fn(pred_keypoints_3d, gt_keypoints_3d)).sum(dim=(1,2))
+        return loss.sum()
+
+class ParameterLoss(nn.Module):
+
+    def __init__(self):
+        """
+        SMPL parameter loss module.
+        """
+        super(ParameterLoss, self).__init__()
+        self.loss_fn = nn.MSELoss(reduction='none')
+
+    def forward(self, pred_param: torch.Tensor, gt_param: torch.Tensor, has_param: torch.Tensor):
+        """
+        Compute SMPL parameter loss.
+        Args:
+            pred_param (torch.Tensor): Tensor of shape [B, S, ...] containing the predicted parameters (body pose / global orientation / betas)
+            gt_param (torch.Tensor): Tensor of shape [B, S, ...] containing the ground truth SMPL parameters.
+        Returns:
+            torch.Tensor: L2 parameter loss loss.
+        """
+        batch_size = pred_param.shape[0]
+        num_dims = len(pred_param.shape)
+        mask_dimension = [batch_size] + [1] * (num_dims-1)
+        has_param = has_param.type(pred_param.type()).view(*mask_dimension)
+        loss_param = (has_param * self.loss_fn(pred_param, gt_param))
+        return loss_param.sum()

hmr2/models/smpl_wrapper.py

@@ -0,0 +1,41 @@
+import torch
+import numpy as np
+import pickle
+from typing import Optional
+import smplx
+from smplx.lbs import vertices2joints
+from smplx.utils import SMPLOutput
+
+
+class SMPL(smplx.SMPLLayer):
+    def __init__(self, *args, joint_regressor_extra: Optional[str] = None, update_hips: bool = False, **kwargs):
+        """
+        Extension of the official SMPL implementation to support more joints.
+        Args:
+            Same as SMPLLayer.
+            joint_regressor_extra (str): Path to extra joint regressor.
+        """
+        super(SMPL, self).__init__(*args, **kwargs)
+        smpl_to_openpose = [24, 12, 17, 19, 21, 16, 18, 20, 0, 2, 5, 8, 1, 4,
+                            7, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34]
+            
+        if joint_regressor_extra is not None:
+            self.register_buffer('joint_regressor_extra', torch.tensor(pickle.load(open(joint_regressor_extra, 'rb'), encoding='latin1'), dtype=torch.float32))
+        self.register_buffer('joint_map', torch.tensor(smpl_to_openpose, dtype=torch.long))
+        self.update_hips = update_hips
+
+    def forward(self, *args, **kwargs) -> SMPLOutput:
+        """
+        Run forward pass. Same as SMPL and also append an extra set of joints if joint_regressor_extra is specified.
+        """
+        smpl_output = super(SMPL, self).forward(*args, **kwargs)
+        joints = smpl_output.joints[:, self.joint_map, :]
+        if self.update_hips:
+            joints[:,[9,12]] = joints[:,[9,12]] + \
+                0.25*(joints[:,[9,12]]-joints[:,[12,9]]) + \
+                0.5*(joints[:,[8]] - 0.5*(joints[:,[9,12]] + joints[:,[12,9]]))
+        if hasattr(self, 'joint_regressor_extra'):
+            extra_joints = vertices2joints(self.joint_regressor_extra, smpl_output.vertices)
+            joints = torch.cat([joints, extra_joints], dim=1)
+        smpl_output.joints = joints
+        return smpl_output

hmr2/utils/__init__.py

@@ -0,0 +1,25 @@
+import torch
+from typing import Any
+
+from .renderer import Renderer
+from .mesh_renderer import MeshRenderer
+from .skeleton_renderer import SkeletonRenderer
+from .pose_utils import eval_pose, Evaluator
+
+def recursive_to(x: Any, target: torch.device):
+    """
+    Recursively transfer a batch of data to the target device
+    Args:
+        x (Any): Batch of data.
+        target (torch.device): Target device.
+    Returns:
+        Batch of data where all tensors are transfered to the target device.
+    """
+    if isinstance(x, dict):
+        return {k: recursive_to(v, target) for k, v in x.items()}
+    elif isinstance(x, torch.Tensor):
+        return x.to(target)
+    elif isinstance(x, list):
+        return [recursive_to(i, target) for i in x]
+    else:
+        return x

hmr2/utils/geometry.py

@@ -0,0 +1,102 @@
+from typing import Optional
+import torch
+from torch.nn import functional as F
+
+def aa_to_rotmat(theta: torch.Tensor):
+    """
+    Convert axis-angle representation to rotation matrix.
+    Works by first converting it to a quaternion.
+    Args:
+        theta (torch.Tensor): Tensor of shape (B, 3) containing axis-angle representations.
+    Returns:
+        torch.Tensor: Corresponding rotation matrices with shape (B, 3, 3).
+    """
+    norm = torch.norm(theta + 1e-8, p = 2, dim = 1)
+    angle = torch.unsqueeze(norm, -1)
+    normalized = torch.div(theta, angle)
+    angle = angle * 0.5
+    v_cos = torch.cos(angle)
+    v_sin = torch.sin(angle)
+    quat = torch.cat([v_cos, v_sin * normalized], dim = 1)
+    return quat_to_rotmat(quat)
+
+def quat_to_rotmat(quat: torch.Tensor) -> torch.Tensor:
+    """
+    Convert quaternion representation to rotation matrix.
+    Args:
+        quat (torch.Tensor) of shape (B, 4); 4 <===> (w, x, y, z).
+    Returns:
+        torch.Tensor: Corresponding rotation matrices with shape (B, 3, 3).
+    """
+    norm_quat = quat
+    norm_quat = norm_quat/norm_quat.norm(p=2, dim=1, keepdim=True)
+    w, x, y, z = norm_quat[:,0], norm_quat[:,1], norm_quat[:,2], norm_quat[:,3]
+
+    B = quat.size(0)
+
+    w2, x2, y2, z2 = w.pow(2), x.pow(2), y.pow(2), z.pow(2)
+    wx, wy, wz = w*x, w*y, w*z
+    xy, xz, yz = x*y, x*z, y*z
+
+    rotMat = torch.stack([w2 + x2 - y2 - z2, 2*xy - 2*wz, 2*wy + 2*xz,
+                          2*wz + 2*xy, w2 - x2 + y2 - z2, 2*yz - 2*wx,
+                          2*xz - 2*wy, 2*wx + 2*yz, w2 - x2 - y2 + z2], dim=1).view(B, 3, 3)
+    return rotMat
+
+
+def rot6d_to_rotmat(x: torch.Tensor) -> torch.Tensor:
+    """
+    Convert 6D rotation representation to 3x3 rotation matrix.
+    Based on Zhou et al., "On the Continuity of Rotation Representations in Neural Networks", CVPR 2019
+    Args:
+        x (torch.Tensor): (B,6) Batch of 6-D rotation representations.
+    Returns:
+        torch.Tensor: Batch of corresponding rotation matrices with shape (B,3,3).
+    """
+    x = x.reshape(-1,2,3).permute(0, 2, 1).contiguous()
+    a1 = x[:, :, 0]
+    a2 = x[:, :, 1]
+    b1 = F.normalize(a1)
+    b2 = F.normalize(a2 - torch.einsum('bi,bi->b', b1, a2).unsqueeze(-1) * b1)
+    b3 = torch.cross(b1, b2)
+    return torch.stack((b1, b2, b3), dim=-1)
+
+def perspective_projection(points: torch.Tensor,
+                           translation: torch.Tensor,
+                           focal_length: torch.Tensor,
+                           camera_center: Optional[torch.Tensor] = None,
+                           rotation: Optional[torch.Tensor] = None) -> torch.Tensor:
+    """
+    Computes the perspective projection of a set of 3D points.
+    Args:
+        points (torch.Tensor): Tensor of shape (B, N, 3) containing the input 3D points.
+        translation (torch.Tensor): Tensor of shape (B, 3) containing the 3D camera translation.
+        focal_length (torch.Tensor): Tensor of shape (B, 2) containing the focal length in pixels.
+        camera_center (torch.Tensor): Tensor of shape (B, 2) containing the camera center in pixels.
+        rotation (torch.Tensor): Tensor of shape (B, 3, 3) containing the camera rotation.
+    Returns:
+        torch.Tensor: Tensor of shape (B, N, 2) containing the projection of the input points.
+    """
+    batch_size = points.shape[0]
+    if rotation is None:
+        rotation = torch.eye(3, device=points.device, dtype=points.dtype).unsqueeze(0).expand(batch_size, -1, -1)
+    if camera_center is None:
+        camera_center = torch.zeros(batch_size, 2, device=points.device, dtype=points.dtype)
+    # Populate intrinsic camera matrix K.
+    K = torch.zeros([batch_size, 3, 3], device=points.device, dtype=points.dtype)
+    K[:,0,0] = focal_length[:,0]
+    K[:,1,1] = focal_length[:,1]
+    K[:,2,2] = 1.
+    K[:,:-1, -1] = camera_center
+
+    # Transform points
+    points = torch.einsum('bij,bkj->bki', rotation, points)
+    points = points + translation.unsqueeze(1)
+
+    # Apply perspective distortion
+    projected_points = points / points[:,:,-1].unsqueeze(-1)
+
+    # Apply camera intrinsics
+    projected_points = torch.einsum('bij,bkj->bki', K, projected_points)
+
+    return projected_points[:, :, :-1]

hmr2/utils/mesh_renderer.py

@@ -0,0 +1,149 @@
+import os
+if 'PYOPENGL_PLATFORM' not in os.environ:
+    os.environ['PYOPENGL_PLATFORM'] = 'egl'
+import torch
+from torchvision.utils import make_grid
+import numpy as np
+import pyrender
+import trimesh
+import cv2
+import torch.nn.functional as F
+
+from .render_openpose import render_openpose
+
+def create_raymond_lights():
+    import pyrender
+    thetas = np.pi * np.array([1.0 / 6.0, 1.0 / 6.0, 1.0 / 6.0])
+    phis = np.pi * np.array([0.0, 2.0 / 3.0, 4.0 / 3.0])
+
+    nodes = []
+
+    for phi, theta in zip(phis, thetas):
+        xp = np.sin(theta) * np.cos(phi)
+        yp = np.sin(theta) * np.sin(phi)
+        zp = np.cos(theta)
+
+        z = np.array([xp, yp, zp])
+        z = z / np.linalg.norm(z)
+        x = np.array([-z[1], z[0], 0.0])
+        if np.linalg.norm(x) == 0:
+            x = np.array([1.0, 0.0, 0.0])
+        x = x / np.linalg.norm(x)
+        y = np.cross(z, x)
+
+        matrix = np.eye(4)
+        matrix[:3,:3] = np.c_[x,y,z]
+        nodes.append(pyrender.Node(
+            light=pyrender.DirectionalLight(color=np.ones(3), intensity=1.0),
+            matrix=matrix
+        ))
+
+    return nodes
+
+class MeshRenderer:
+
+    def __init__(self, cfg, faces=None):
+        self.cfg = cfg
+        self.focal_length = cfg.EXTRA.FOCAL_LENGTH
+        self.img_res = cfg.MODEL.IMAGE_SIZE
+        self.renderer = pyrender.OffscreenRenderer(viewport_width=self.img_res,
+                                       viewport_height=self.img_res,
+                                       point_size=1.0)
+        
+        self.camera_center = [self.img_res // 2, self.img_res // 2]
+        self.faces = faces
+
+    def visualize(self, vertices, camera_translation, images, focal_length=None, nrow=3, padding=2):
+        images_np = np.transpose(images, (0,2,3,1))
+        rend_imgs = []
+        for i in range(vertices.shape[0]):
+            fl = self.focal_length
+            rend_img = torch.from_numpy(np.transpose(self.__call__(vertices[i], camera_translation[i], images_np[i], focal_length=fl, side_view=False), (2,0,1))).float()
+            rend_img_side = torch.from_numpy(np.transpose(self.__call__(vertices[i], camera_translation[i], images_np[i], focal_length=fl, side_view=True), (2,0,1))).float()
+            rend_imgs.append(torch.from_numpy(images[i]))
+            rend_imgs.append(rend_img)
+            rend_imgs.append(rend_img_side)
+        rend_imgs = make_grid(rend_imgs, nrow=nrow, padding=padding)
+        return rend_imgs
+
+    def visualize_tensorboard(self, vertices, camera_translation, images, pred_keypoints, gt_keypoints, focal_length=None, nrow=5, padding=2):
+        images_np = np.transpose(images, (0,2,3,1))
+        rend_imgs = []
+        pred_keypoints = np.concatenate((pred_keypoints, np.ones_like(pred_keypoints)[:, :, [0]]), axis=-1)
+        pred_keypoints = self.img_res * (pred_keypoints + 0.5)
+        gt_keypoints[:, :, :-1] = self.img_res * (gt_keypoints[:, :, :-1] + 0.5)
+        keypoint_matches = [(1, 12), (2, 8), (3, 7), (4, 6), (5, 9), (6, 10), (7, 11), (8, 14), (9, 2), (10, 1), (11, 0), (12, 3), (13, 4), (14, 5)]
+        for i in range(vertices.shape[0]):
+            fl = self.focal_length
+            rend_img = torch.from_numpy(np.transpose(self.__call__(vertices[i], camera_translation[i], images_np[i], focal_length=fl, side_view=False), (2,0,1))).float()
+            rend_img_side = torch.from_numpy(np.transpose(self.__call__(vertices[i], camera_translation[i], images_np[i], focal_length=fl, side_view=True), (2,0,1))).float()
+            body_keypoints = pred_keypoints[i, :25]
+            extra_keypoints = pred_keypoints[i, -19:]
+            for pair in keypoint_matches:
+                body_keypoints[pair[0], :] = extra_keypoints[pair[1], :]
+            pred_keypoints_img = render_openpose(255 * images_np[i].copy(), body_keypoints) / 255
+            body_keypoints = gt_keypoints[i, :25]
+            extra_keypoints = gt_keypoints[i, -19:]
+            for pair in keypoint_matches:
+                if extra_keypoints[pair[1], -1] > 0 and body_keypoints[pair[0], -1] == 0:
+                    body_keypoints[pair[0], :] = extra_keypoints[pair[1], :]
+            gt_keypoints_img = render_openpose(255*images_np[i].copy(), body_keypoints) / 255
+            rend_imgs.append(torch.from_numpy(images[i]))
+            rend_imgs.append(rend_img)
+            rend_imgs.append(rend_img_side)
+            rend_imgs.append(torch.from_numpy(pred_keypoints_img).permute(2,0,1))
+            rend_imgs.append(torch.from_numpy(gt_keypoints_img).permute(2,0,1))
+        rend_imgs = make_grid(rend_imgs, nrow=nrow, padding=padding)
+        return rend_imgs
+
+    def __call__(self, vertices, camera_translation, image, focal_length=5000, text=None, resize=None, side_view=False, baseColorFactor=(1.0, 1.0, 0.9, 1.0), rot_angle=90):
+        renderer = pyrender.OffscreenRenderer(viewport_width=image.shape[1],
+                                              viewport_height=image.shape[0],
+                                              point_size=1.0)
+        material = pyrender.MetallicRoughnessMaterial(
+            metallicFactor=0.0,
+            alphaMode='OPAQUE',
+            baseColorFactor=baseColorFactor)
+
+        camera_translation[0] *= -1.
+
+        mesh = trimesh.Trimesh(vertices.copy(), self.faces.copy())
+        if side_view:
+            rot = trimesh.transformations.rotation_matrix(
+                np.radians(rot_angle), [0, 1, 0])
+            mesh.apply_transform(rot)
+        rot = trimesh.transformations.rotation_matrix(
+            np.radians(180), [1, 0, 0])
+        mesh.apply_transform(rot)
+        mesh = pyrender.Mesh.from_trimesh(mesh, material=material)
+
+        scene = pyrender.Scene(bg_color=[0.0, 0.0, 0.0, 0.0],
+                               ambient_light=(0.3, 0.3, 0.3))
+        scene.add(mesh, 'mesh')
+
+        camera_pose = np.eye(4)
+        camera_pose[:3, 3] = camera_translation
+        camera_center = [image.shape[1] / 2., image.shape[0] / 2.]
+        camera = pyrender.IntrinsicsCamera(fx=focal_length, fy=focal_length,
+                                           cx=camera_center[0], cy=camera_center[1])
+        scene.add(camera, pose=camera_pose)
+
+
+        light_nodes = create_raymond_lights()
+        for node in light_nodes:
+            scene.add_node(node)
+
+        color, rend_depth = renderer.render(scene, flags=pyrender.RenderFlags.RGBA)
+        color = color.astype(np.float32) / 255.0
+        valid_mask = (color[:, :, -1] > 0)[:, :, np.newaxis]
+        if not side_view:
+            output_img = (color[:, :, :3] * valid_mask +
+                      (1 - valid_mask) * image)
+        else:
+            output_img = color[:, :, :3]
+        if resize is not None:
+            output_img = cv2.resize(output_img, resize)
+
+        output_img = output_img.astype(np.float32)
+        renderer.delete()
+        return output_img

hmr2/utils/pose_utils.py

@@ -0,0 +1,306 @@
+"""
+Code adapted from: https://github.com/akanazawa/hmr/blob/master/src/benchmark/eval_util.py
+"""
+
+import torch
+import numpy as np
+from typing import Optional, Dict, List, Tuple
+
+def compute_similarity_transform(S1: torch.Tensor, S2: torch.Tensor) -> torch.Tensor:
+    """
+    Computes a similarity transform (sR, t) in a batched way that takes
+    a set of 3D points S1 (B, N, 3) closest to a set of 3D points S2 (B, N, 3),
+    where R is a 3x3 rotation matrix, t 3x1 translation, s scale.
+    i.e. solves the orthogonal Procrutes problem.
+    Args:
+        S1 (torch.Tensor): First set of points of shape (B, N, 3).
+        S2 (torch.Tensor): Second set of points of shape (B, N, 3).
+    Returns:
+        (torch.Tensor): The first set of points after applying the similarity transformation.
+    """
+
+    batch_size = S1.shape[0]
+    S1 = S1.permute(0, 2, 1)
+    S2 = S2.permute(0, 2, 1)
+    # 1. Remove mean.
+    mu1 = S1.mean(dim=2, keepdim=True)
+    mu2 = S2.mean(dim=2, keepdim=True)
+    X1 = S1 - mu1
+    X2 = S2 - mu2
+
+    # 2. Compute variance of X1 used for scale.
+    var1 = (X1**2).sum(dim=(1,2))
+
+    # 3. The outer product of X1 and X2.
+    K = torch.matmul(X1, X2.permute(0, 2, 1))
+
+    # 4. Solution that Maximizes trace(R'K) is R=U*V', where U, V are singular vectors of K.
+    U, s, V = torch.svd(K)
+    Vh = V.permute(0, 2, 1)
+
+    # Construct Z that fixes the orientation of R to get det(R)=1.
+    Z = torch.eye(U.shape[1], device=U.device).unsqueeze(0).repeat(batch_size, 1, 1)
+    Z[:, -1, -1] *= torch.sign(torch.linalg.det(torch.matmul(U, Vh)))
+
+    # Construct R.
+    R = torch.matmul(torch.matmul(V, Z), U.permute(0, 2, 1))
+
+    # 5. Recover scale.
+    trace = torch.matmul(R, K).diagonal(offset=0, dim1=-1, dim2=-2).sum(dim=-1)
+    scale = (trace / var1).unsqueeze(dim=-1).unsqueeze(dim=-1)
+
+    # 6. Recover translation.
+    t = mu2 - scale*torch.matmul(R, mu1)
+
+    # 7. Error:
+    S1_hat = scale*torch.matmul(R, S1) + t
+
+    return S1_hat.permute(0, 2, 1)
+
+def reconstruction_error(S1, S2) -> np.array:
+    """
+    Computes the mean Euclidean distance of 2 set of points S1, S2 after performing Procrustes alignment.
+    Args:
+        S1 (torch.Tensor): First set of points of shape (B, N, 3).
+        S2 (torch.Tensor): Second set of points of shape (B, N, 3).
+    Returns:
+        (np.array): Reconstruction error.
+    """
+    S1_hat = compute_similarity_transform(S1, S2)
+    re = torch.sqrt( ((S1_hat - S2)** 2).sum(dim=-1)).mean(dim=-1)
+    return re
+
+def eval_pose(pred_joints, gt_joints) -> Tuple[np.array, np.array]:
+    """
+    Compute joint errors in mm before and after Procrustes alignment.
+    Args:
+        pred_joints (torch.Tensor): Predicted 3D joints of shape (B, N, 3).
+        gt_joints (torch.Tensor): Ground truth 3D joints of shape (B, N, 3).
+    Returns:
+        Tuple[np.array, np.array]: Joint errors in mm before and after alignment.
+    """
+    # Absolute error (MPJPE)
+    mpjpe = torch.sqrt(((pred_joints - gt_joints) ** 2).sum(dim=-1)).mean(dim=-1).cpu().numpy()
+
+    # Reconstuction_error
+    r_error = reconstruction_error(pred_joints, gt_joints).cpu().numpy()
+    return 1000 * mpjpe, 1000 * r_error
+
+class Evaluator:
+
+    def __init__(self,
+                 dataset_length: int,
+                 keypoint_list: List,
+                 pelvis_ind: int,
+                 metrics: List = ['mode_mpjpe', 'mode_re', 'min_mpjpe', 'min_re'],
+                 pck_thresholds: Optional[List] = None):
+        """
+        Class used for evaluating trained models on different 3D pose datasets.
+        Args:
+            dataset_length (int): Total dataset length.
+            keypoint_list [List]: List of keypoints used for evaluation.
+            pelvis_ind (int): Index of pelvis keypoint; used for aligning the predictions and ground truth.
+            metrics [List]: List of evaluation metrics to record.
+        """
+        self.dataset_length = dataset_length
+        self.keypoint_list = keypoint_list
+        self.pelvis_ind = pelvis_ind
+        self.metrics = metrics
+        for metric in self.metrics:
+            setattr(self, metric, np.zeros((dataset_length,)))
+        self.counter = 0
+        if pck_thresholds is None:
+            self.pck_evaluator = None
+        else:
+            self.pck_evaluator = EvaluatorPCK(pck_thresholds)
+
+    def log(self):
+        """
+        Print current evaluation metrics
+        """
+        if self.counter == 0:
+            print('Evaluation has not started')
+            return
+        print(f'{self.counter} / {self.dataset_length} samples')
+        if self.pck_evaluator is not None:
+            self.pck_evaluator.log()
+        for metric in self.metrics:
+            if metric in ['mode_mpjpe', 'mode_re', 'min_mpjpe', 'min_re']:
+                unit = 'mm'
+            else:
+                unit = ''
+            print(f'{metric}: {getattr(self, metric)[:self.counter].mean()} {unit}')
+        print('***')
+
+    def get_metrics_dict(self) -> Dict:
+        """
+        Returns:
+            Dict: Dictionary of evaluation metrics.
+        """
+        d1 = {metric: getattr(self, metric)[:self.counter].mean() for metric in self.metrics}
+        if self.pck_evaluator is not None:
+            d2 = self.pck_evaluator.get_metrics_dict()
+            d1.update(d2)
+        return d1
+
+    def __call__(self, output: Dict, batch: Dict, opt_output: Optional[Dict] = None):
+        """
+        Evaluate current batch.
+        Args:
+            output (Dict): Regression output.
+            batch (Dict): Dictionary containing images and their corresponding annotations.
+            opt_output (Dict): Optimization output.
+        """
+        if self.pck_evaluator is not None:
+            self.pck_evaluator(output, batch, opt_output)
+
+        pred_keypoints_3d = output['pred_keypoints_3d'].detach()
+        pred_keypoints_3d = pred_keypoints_3d[:,None,:,:]
+        batch_size = pred_keypoints_3d.shape[0]
+        num_samples = pred_keypoints_3d.shape[1]
+        gt_keypoints_3d = batch['keypoints_3d'][:, :, :-1].unsqueeze(1).repeat(1, num_samples, 1, 1)
+
+        # Align predictions and ground truth such that the pelvis location is at the origin
+        pred_keypoints_3d -= pred_keypoints_3d[:, :, [self.pelvis_ind]]
+        gt_keypoints_3d -= gt_keypoints_3d[:, :, [self.pelvis_ind]]
+
+        # Compute joint errors
+        mpjpe, re = eval_pose(pred_keypoints_3d.reshape(batch_size * num_samples, -1, 3)[:, self.keypoint_list], gt_keypoints_3d.reshape(batch_size * num_samples, -1 ,3)[:, self.keypoint_list])
+        mpjpe = mpjpe.reshape(batch_size, num_samples)
+        re = re.reshape(batch_size, num_samples)
+
+        # Compute 2d keypoint errors
+        pred_keypoints_2d = output['pred_keypoints_2d'].detach()
+        pred_keypoints_2d = pred_keypoints_2d[:,None,:,:]
+        gt_keypoints_2d = batch['keypoints_2d'][:,None,:,:].repeat(1, num_samples, 1, 1)
+        conf = gt_keypoints_2d[:, :, :, -1].clone()
+        kp_err = torch.nn.functional.mse_loss(
+                        pred_keypoints_2d,
+                        gt_keypoints_2d[:, :, :, :-1],
+                        reduction='none'
+                    ).sum(dim=3)
+        kp_l2_loss = (conf * kp_err).mean(dim=2)
+        kp_l2_loss = kp_l2_loss.detach().cpu().numpy()
+
+        # Compute joint errors after optimization, if available.
+        if opt_output is not None:
+            opt_keypoints_3d = opt_output['model_joints']
+            opt_keypoints_3d -= opt_keypoints_3d[:, [self.pelvis_ind]]
+            opt_mpjpe, opt_re = eval_pose(opt_keypoints_3d[:, self.keypoint_list], gt_keypoints_3d[:, 0, self.keypoint_list])
+
+        # The 0-th sample always corresponds to the mode
+        if hasattr(self, 'mode_mpjpe'):
+            mode_mpjpe = mpjpe[:, 0]
+            self.mode_mpjpe[self.counter:self.counter+batch_size] = mode_mpjpe
+        if hasattr(self, 'mode_re'):
+            mode_re = re[:, 0]
+            self.mode_re[self.counter:self.counter+batch_size] = mode_re
+        if hasattr(self, 'mode_kpl2'):
+            mode_kpl2 = kp_l2_loss[:, 0]
+            self.mode_kpl2[self.counter:self.counter+batch_size] = mode_kpl2
+        if hasattr(self, 'min_mpjpe'):
+            min_mpjpe = mpjpe.min(axis=-1)
+            self.min_mpjpe[self.counter:self.counter+batch_size] = min_mpjpe
+        if hasattr(self, 'min_re'):
+            min_re = re.min(axis=-1)
+            self.min_re[self.counter:self.counter+batch_size] = min_re
+        if hasattr(self, 'min_kpl2'):
+            min_kpl2 = kp_l2_loss.min(axis=-1)
+            self.min_kpl2[self.counter:self.counter+batch_size] = min_kpl2
+        if hasattr(self, 'opt_mpjpe'):
+            self.opt_mpjpe[self.counter:self.counter+batch_size] = opt_mpjpe
+        if hasattr(self, 'opt_re'):
+            self.opt_re[self.counter:self.counter+batch_size] = opt_re
+
+        self.counter += batch_size
+
+        if hasattr(self, 'mode_mpjpe') and hasattr(self, 'mode_re'):
+            return {
+                'mode_mpjpe': mode_mpjpe,
+                'mode_re': mode_re,
+            }
+        else:
+            return {}
+
+
+class EvaluatorPCK:
+
+    def __init__(self, thresholds: List = [0.05, 0.1, 0.2, 0.3, 0.4, 0.5],):
+        """
+        Class used for evaluating trained models on different 3D pose datasets.
+        Args:
+            thresholds [List]: List of PCK thresholds to evaluate.
+            metrics [List]: List of evaluation metrics to record.
+        """
+        self.thresholds = thresholds
+        self.pred_kp_2d = []
+        self.gt_kp_2d = []
+        self.gt_conf_2d = []
+        self.counter = 0
+
+    def log(self):
+        """
+        Print current evaluation metrics
+        """
+        if self.counter == 0:
+            print('Evaluation has not started')
+            return
+        print(f'{self.counter} samples')
+        metrics_dict = self.get_metrics_dict()
+        for metric in metrics_dict:
+            print(f'{metric}: {metrics_dict[metric]}')
+        print('***')
+
+    def get_metrics_dict(self) -> Dict:
+        """
+        Returns:
+            Dict: Dictionary of evaluation metrics.
+        """
+        pcks = self.compute_pcks()
+        metrics = {}
+        for thr, (acc,avg_acc,cnt) in zip(self.thresholds, pcks):
+            metrics.update({f'kp{i}_pck_{thr}': float(a) for i, a in enumerate(acc) if a>=0})
+            metrics.update({f'kpAvg_pck_{thr}': float(avg_acc)})
+        return metrics
+
+    def compute_pcks(self):
+        pred_kp_2d = np.concatenate(self.pred_kp_2d, axis=0)
+        gt_kp_2d = np.concatenate(self.gt_kp_2d, axis=0)
+        gt_conf_2d = np.concatenate(self.gt_conf_2d, axis=0)
+        assert pred_kp_2d.shape == gt_kp_2d.shape
+        assert pred_kp_2d[..., 0].shape == gt_conf_2d.shape
+        assert pred_kp_2d.shape[1] == 1 # num_samples
+
+        from mmpose.core.evaluation import keypoint_pck_accuracy
+        pcks = [
+            keypoint_pck_accuracy(
+                pred_kp_2d[:, 0, :, :],
+                gt_kp_2d[:, 0, :, :],
+                gt_conf_2d[:, 0, :]>0.5,
+                thr=thr,
+                normalize = np.ones((len(pred_kp_2d),2))   # Already in [-0.5,0.5] range. No need to normalize
+            )
+            for thr in self.thresholds
+        ]
+        return pcks
+
+    def __call__(self, output: Dict, batch: Dict, opt_output: Optional[Dict] = None):
+        """
+        Evaluate current batch.
+        Args:
+            output (Dict): Regression output.
+            batch (Dict): Dictionary containing images and their corresponding annotations.
+            opt_output (Dict): Optimization output.
+        """
+        pred_keypoints_2d = output['pred_keypoints_2d'].detach()
+        num_samples = 1
+        batch_size = pred_keypoints_2d.shape[0]
+
+        pred_keypoints_2d = pred_keypoints_2d[:,None,:,:]
+        gt_keypoints_2d = batch['keypoints_2d'][:,None,:,:].repeat(1, num_samples, 1, 1)
+
+        self.pred_kp_2d.append(pred_keypoints_2d[:, :, :, :2].detach().cpu().numpy())
+        self.gt_conf_2d.append(gt_keypoints_2d[:, :, :, -1].detach().cpu().numpy())
+        self.gt_kp_2d.append(gt_keypoints_2d[:, :, :, :2].detach().cpu().numpy())
+
+        self.counter += batch_size

hmr2/utils/render_openpose.py

@@ -0,0 +1,149 @@
+"""
+Render OpenPose keypoints.
+Code was ported to Python from the official C++ implementation https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/src/openpose/utilities/keypoint.cpp
+"""
+import cv2
+import math
+import numpy as np
+from typing import List, Tuple
+
+def get_keypoints_rectangle(keypoints: np.array, threshold: float) -> Tuple[float, float, float]:
+    """
+    Compute rectangle enclosing keypoints above the threshold.
+    Args:
+        keypoints (np.array): Keypoint array of shape (N, 3).
+        threshold (float): Confidence visualization threshold.
+    Returns:
+        Tuple[float, float, float]: Rectangle width, height and area.
+    """
+    valid_ind = keypoints[:, -1] > threshold
+    if valid_ind.sum() > 0:
+        valid_keypoints = keypoints[valid_ind][:, :-1]
+        max_x = valid_keypoints[:,0].max()
+        max_y = valid_keypoints[:,1].max()
+        min_x = valid_keypoints[:,0].min()
+        min_y = valid_keypoints[:,1].min()
+        width = max_x - min_x
+        height = max_y - min_y
+        area = width * height
+        return width, height, area
+    else:
+        return 0,0,0
+
+def render_keypoints(img: np.array,
+                     keypoints: np.array,
+                     pairs: List,
+                     colors: List,
+                     thickness_circle_ratio: float,
+                     thickness_line_ratio_wrt_circle: float,
+                     pose_scales: List,
+                     threshold: float = 0.1) -> np.array:
+    """
+    Render keypoints on input image.
+    Args:
+        img (np.array): Input image of shape (H, W, 3) with pixel values in the [0,255] range.
+        keypoints (np.array): Keypoint array of shape (N, 3).
+        pairs (List): List of keypoint pairs per limb.
+        colors: (List): List of colors per keypoint.
+        thickness_circle_ratio (float): Circle thickness ratio.
+        thickness_line_ratio_wrt_circle (float): Line thickness ratio wrt the circle.
+        pose_scales (List): List of pose scales.
+        threshold (float): Only visualize keypoints with confidence above the threshold.
+    Returns:
+        (np.array): Image of shape (H, W, 3) with keypoints drawn on top of the original image. 
+    """
+    img_orig = img.copy()
+    width, height = img.shape[1], img.shape[2]
+    area = width * height
+
+    lineType = 8
+    shift = 0
+    numberColors = len(colors)
+    thresholdRectangle = 0.1
+
+    person_width, person_height, person_area = get_keypoints_rectangle(keypoints, thresholdRectangle)
+    if person_area > 0:
+        ratioAreas = min(1, max(person_width / width, person_height / height))
+        thicknessRatio = np.maximum(np.round(math.sqrt(area) * thickness_circle_ratio * ratioAreas), 2)
+        thicknessCircle = np.maximum(1, thicknessRatio if ratioAreas > 0.05 else -np.ones_like(thicknessRatio))
+        thicknessLine = np.maximum(1, np.round(thicknessRatio * thickness_line_ratio_wrt_circle))
+        radius = thicknessRatio / 2
+
+        img = np.ascontiguousarray(img.copy())
+        for i, pair in enumerate(pairs):
+            index1, index2 = pair
+            if keypoints[index1, -1] > threshold and keypoints[index2, -1] > threshold:
+                thicknessLineScaled = int(round(min(thicknessLine[index1], thicknessLine[index2]) * pose_scales[0]))
+                colorIndex = index2
+                color = colors[colorIndex % numberColors]
+                keypoint1 = keypoints[index1, :-1].astype(np.int)
+                keypoint2 = keypoints[index2, :-1].astype(np.int)
+                cv2.line(img, tuple(keypoint1.tolist()), tuple(keypoint2.tolist()), tuple(color.tolist()), thicknessLineScaled, lineType, shift)
+        for part in range(len(keypoints)):
+            faceIndex = part
+            if keypoints[faceIndex, -1] > threshold:
+                radiusScaled = int(round(radius[faceIndex] * pose_scales[0]))
+                thicknessCircleScaled = int(round(thicknessCircle[faceIndex] * pose_scales[0]))
+                colorIndex = part
+                color = colors[colorIndex % numberColors]
+                center = keypoints[faceIndex, :-1].astype(np.int)
+                cv2.circle(img, tuple(center.tolist()), radiusScaled, tuple(color.tolist()), thicknessCircleScaled, lineType, shift)
+    return img
+
+def render_body_keypoints(img: np.array,
+                          body_keypoints: np.array) -> np.array:
+    """
+    Render OpenPose body keypoints on input image.
+    Args:
+        img (np.array): Input image of shape (H, W, 3) with pixel values in the [0,255] range.
+        body_keypoints (np.array): Keypoint array of shape (N, 3); 3 <====> (x, y, confidence).
+    Returns:
+        (np.array): Image of shape (H, W, 3) with keypoints drawn on top of the original image. 
+    """
+
+    thickness_circle_ratio = 1./75. * np.ones(body_keypoints.shape[0])
+    thickness_line_ratio_wrt_circle = 0.75
+    pairs = []
+    pairs = [1,8,1,2,1,5,2,3,3,4,5,6,6,7,8,9,9,10,10,11,8,12,12,13,13,14,1,0,0,15,15,17,0,16,16,18,14,19,19,20,14,21,11,22,22,23,11,24]
+    pairs = np.array(pairs).reshape(-1,2)
+    colors = [255.,     0.,     85.,
+              255.,     0.,     0.,
+              255.,    85.,     0.,
+              255.,   170.,     0.,
+              255.,   255.,     0.,
+              170.,   255.,     0.,
+               85.,   255.,     0.,
+                0.,   255.,     0.,
+              255.,     0.,     0.,
+                0.,   255.,    85.,
+                0.,   255.,   170.,
+                0.,   255.,   255.,
+                0.,   170.,   255.,
+                0.,    85.,   255.,
+                0.,     0.,   255.,
+              255.,     0.,   170.,
+              170.,     0.,   255.,
+              255.,     0.,   255.,
+               85.,     0.,   255.,
+                0.,     0.,   255.,
+                0.,     0.,   255.,
+                0.,     0.,   255.,
+                0.,   255.,   255.,
+                0.,   255.,   255.,
+                0.,   255.,   255.]
+    colors = np.array(colors).reshape(-1,3)
+    pose_scales = [1]
+    return render_keypoints(img, body_keypoints, pairs, colors, thickness_circle_ratio, thickness_line_ratio_wrt_circle, pose_scales, 0.1)
+
+def render_openpose(img: np.array,
+                    body_keypoints: np.array) -> np.array:
+    """
+    Render keypoints in the OpenPose format on input image.
+    Args:
+        img (np.array): Input image of shape (H, W, 3) with pixel values in the [0,255] range.
+        body_keypoints (np.array): Keypoint array of shape (N, 3); 3 <====> (x, y, confidence).
+    Returns:
+        (np.array): Image of shape (H, W, 3) with keypoints drawn on top of the original image. 
+    """
+    img = render_body_keypoints(img, body_keypoints)
+    return img

hmr2/utils/renderer.py

@@ -0,0 +1,396 @@
+import os
+if 'PYOPENGL_PLATFORM' not in os.environ:
+    os.environ['PYOPENGL_PLATFORM'] = 'egl'
+import torch
+import numpy as np
+import pyrender
+import trimesh
+import cv2
+from yacs.config import CfgNode
+from typing import List, Optional
+
+def cam_crop_to_full(cam_bbox, box_center, box_size, img_size, focal_length=5000.):
+    # Convert cam_bbox to full image
+    img_w, img_h = img_size[:, 0], img_size[:, 1]
+    cx, cy, b = box_center[:, 0], box_center[:, 1], box_size
+    w_2, h_2 = img_w / 2., img_h / 2.
+    bs = b * cam_bbox[:, 0] + 1e-9
+    tz = 2 * focal_length / bs
+    tx = (2 * (cx - w_2) / bs) + cam_bbox[:, 1]
+    ty = (2 * (cy - h_2) / bs) + cam_bbox[:, 2]
+    full_cam = torch.stack([tx, ty, tz], dim=-1)
+    return full_cam
+
+def get_light_poses(n_lights=5, elevation=np.pi / 3, dist=12):
+    # get lights in a circle around origin at elevation
+    thetas = elevation * np.ones(n_lights)
+    phis = 2 * np.pi * np.arange(n_lights) / n_lights
+    poses = []
+    trans = make_translation(torch.tensor([0, 0, dist]))
+    for phi, theta in zip(phis, thetas):
+        rot = make_rotation(rx=-theta, ry=phi, order="xyz")
+        poses.append((rot @ trans).numpy())
+    return poses
+
+def make_translation(t):
+    return make_4x4_pose(torch.eye(3), t)
+
+def make_rotation(rx=0, ry=0, rz=0, order="xyz"):
+    Rx = rotx(rx)
+    Ry = roty(ry)
+    Rz = rotz(rz)
+    if order == "xyz":
+        R = Rz @ Ry @ Rx
+    elif order == "xzy":
+        R = Ry @ Rz @ Rx
+    elif order == "yxz":
+        R = Rz @ Rx @ Ry
+    elif order == "yzx":
+        R = Rx @ Rz @ Ry
+    elif order == "zyx":
+        R = Rx @ Ry @ Rz
+    elif order == "zxy":
+        R = Ry @ Rx @ Rz
+    return make_4x4_pose(R, torch.zeros(3))
+
+def make_4x4_pose(R, t):
+    """
+    :param R (*, 3, 3)
+    :param t (*, 3)
+    return (*, 4, 4)
+    """
+    dims = R.shape[:-2]
+    pose_3x4 = torch.cat([R, t.view(*dims, 3, 1)], dim=-1)
+    bottom = (
+        torch.tensor([0, 0, 0, 1], device=R.device)
+        .reshape(*(1,) * len(dims), 1, 4)
+        .expand(*dims, 1, 4)
+    )
+    return torch.cat([pose_3x4, bottom], dim=-2)
+
+
+def rotx(theta):
+    return torch.tensor(
+        [
+            [1, 0, 0],
+            [0, np.cos(theta), -np.sin(theta)],
+            [0, np.sin(theta), np.cos(theta)],
+        ],
+        dtype=torch.float32,
+    )
+
+
+def roty(theta):
+    return torch.tensor(
+        [
+            [np.cos(theta), 0, np.sin(theta)],
+            [0, 1, 0],
+            [-np.sin(theta), 0, np.cos(theta)],
+        ],
+        dtype=torch.float32,
+    )
+
+
+def rotz(theta):
+    return torch.tensor(
+        [
+            [np.cos(theta), -np.sin(theta), 0],
+            [np.sin(theta), np.cos(theta), 0],
+            [0, 0, 1],
+        ],
+        dtype=torch.float32,
+    )
+    
+
+def create_raymond_lights() -> List[pyrender.Node]:
+    """
+    Return raymond light nodes for the scene.
+    """
+    thetas = np.pi * np.array([1.0 / 6.0, 1.0 / 6.0, 1.0 / 6.0])
+    phis = np.pi * np.array([0.0, 2.0 / 3.0, 4.0 / 3.0])
+
+    nodes = []
+
+    for phi, theta in zip(phis, thetas):
+        xp = np.sin(theta) * np.cos(phi)
+        yp = np.sin(theta) * np.sin(phi)
+        zp = np.cos(theta)
+
+        z = np.array([xp, yp, zp])
+        z = z / np.linalg.norm(z)
+        x = np.array([-z[1], z[0], 0.0])
+        if np.linalg.norm(x) == 0:
+            x = np.array([1.0, 0.0, 0.0])
+        x = x / np.linalg.norm(x)
+        y = np.cross(z, x)
+
+        matrix = np.eye(4)
+        matrix[:3,:3] = np.c_[x,y,z]
+        nodes.append(pyrender.Node(
+            light=pyrender.DirectionalLight(color=np.ones(3), intensity=1.0),
+            matrix=matrix
+        ))
+
+    return nodes
+
+class Renderer:
+
+    def __init__(self, cfg: CfgNode, faces: np.array):
+        """
+        Wrapper around the pyrender renderer to render SMPL meshes.
+        Args:
+            cfg (CfgNode): Model config file.
+            faces (np.array): Array of shape (F, 3) containing the mesh faces.
+        """
+        self.cfg = cfg
+        self.focal_length = cfg.EXTRA.FOCAL_LENGTH
+        self.img_res = cfg.MODEL.IMAGE_SIZE
+        # self.renderer = pyrender.OffscreenRenderer(viewport_width=self.img_res,
+        #                                viewport_height=self.img_res,
+        #                                point_size=1.0)
+
+        self.camera_center = [self.img_res // 2, self.img_res // 2]
+        self.faces = faces
+
+    def __call__(self,
+                vertices: np.array,
+                camera_translation: np.array,
+                image: torch.Tensor,
+                full_frame: bool = False,
+                imgname: Optional[str] = None,
+                side_view=False, rot_angle=90,
+                mesh_base_color=(1.0, 1.0, 0.9),
+                scene_bg_color=(0,0,0),
+                return_rgba=False,
+                ) -> np.array:
+        """
+        Render meshes on input image
+        Args:
+            vertices (np.array): Array of shape (V, 3) containing the mesh vertices.
+            camera_translation (np.array): Array of shape (3,) with the camera translation.
+            image (torch.Tensor): Tensor of shape (3, H, W) containing the image crop with normalized pixel values.
+            full_frame (bool): If True, then render on the full image.
+            imgname (Optional[str]): Contains the original image filenamee. Used only if full_frame == True.
+        """
+        
+        if full_frame:
+            image = cv2.imread(imgname).astype(np.float32)[:, :, ::-1] / 255.
+        else:
+            image = image.clone() * torch.tensor(self.cfg.MODEL.IMAGE_STD, device=image.device).reshape(3,1,1)
+            image = image + torch.tensor(self.cfg.MODEL.IMAGE_MEAN, device=image.device).reshape(3,1,1)
+            image = image.permute(1, 2, 0).cpu().numpy()
+
+        renderer = pyrender.OffscreenRenderer(viewport_width=image.shape[1],
+                                              viewport_height=image.shape[0],
+                                              point_size=1.0)
+        material = pyrender.MetallicRoughnessMaterial(
+            metallicFactor=0.0,
+            alphaMode='OPAQUE',
+            baseColorFactor=(*mesh_base_color, 1.0))
+
+        camera_translation[0] *= -1.
+
+        mesh = trimesh.Trimesh(vertices.copy(), self.faces.copy())
+        if side_view:
+            rot = trimesh.transformations.rotation_matrix(
+                np.radians(rot_angle), [0, 1, 0])
+            mesh.apply_transform(rot)
+        rot = trimesh.transformations.rotation_matrix(
+            np.radians(180), [1, 0, 0])
+        mesh.apply_transform(rot)
+        mesh = pyrender.Mesh.from_trimesh(mesh, material=material)
+
+        scene = pyrender.Scene(bg_color=[*scene_bg_color, 0.0],
+                               ambient_light=(0.3, 0.3, 0.3))
+        scene.add(mesh, 'mesh')
+
+        camera_pose = np.eye(4)
+        camera_pose[:3, 3] = camera_translation
+        camera_center = [image.shape[1] / 2., image.shape[0] / 2.]
+        camera = pyrender.IntrinsicsCamera(fx=self.focal_length, fy=self.focal_length,
+                                           cx=camera_center[0], cy=camera_center[1])
+        scene.add(camera, pose=camera_pose)
+
+
+        light_nodes = create_raymond_lights()
+        for node in light_nodes:
+            scene.add_node(node)
+
+        color, rend_depth = renderer.render(scene, flags=pyrender.RenderFlags.RGBA)
+        color = color.astype(np.float32) / 255.0
+        renderer.delete()
+
+        if return_rgba:
+            return color
+
+        valid_mask = (color[:, :, -1])[:, :, np.newaxis]
+        if not side_view:
+            output_img = (color[:, :, :3] * valid_mask + (1 - valid_mask) * image)
+        else:
+            output_img = color[:, :, :3]
+
+        output_img = output_img.astype(np.float32)
+        return output_img
+
+    def vertices_to_trimesh(self, vertices, camera_translation, mesh_base_color=(1.0, 1.0, 0.9), 
+                            rot_axis=[1,0,0], rot_angle=0,):
+        # material = pyrender.MetallicRoughnessMaterial(
+        #     metallicFactor=0.0,
+        #     alphaMode='OPAQUE',
+        #     baseColorFactor=(*mesh_base_color, 1.0))
+        vertex_colors = np.array([(*mesh_base_color, 1.0)] * vertices.shape[0])
+        print(vertices.shape, camera_translation.shape)
+        mesh = trimesh.Trimesh(vertices.copy() + camera_translation, self.faces.copy(), vertex_colors=vertex_colors)
+        # mesh = trimesh.Trimesh(vertices.copy(), self.faces.copy())
+        
+        rot = trimesh.transformations.rotation_matrix(
+                np.radians(rot_angle), rot_axis)
+        mesh.apply_transform(rot)
+
+        rot = trimesh.transformations.rotation_matrix(
+            np.radians(180), [1, 0, 0])
+        mesh.apply_transform(rot)
+        return mesh
+
+    def render_rgba(
+            self,
+            vertices: np.array,
+            cam_t = None,
+            rot=None,
+            rot_axis=[1,0,0],
+            rot_angle=0,
+            camera_z=3,
+            # camera_translation: np.array,
+            mesh_base_color=(1.0, 1.0, 0.9),
+            scene_bg_color=(0,0,0),
+            render_res=[256, 256],
+        ):
+
+        renderer = pyrender.OffscreenRenderer(viewport_width=render_res[0],
+                                              viewport_height=render_res[1],
+                                              point_size=1.0)
+        # material = pyrender.MetallicRoughnessMaterial(
+        #     metallicFactor=0.0,
+        #     alphaMode='OPAQUE',
+        #     baseColorFactor=(*mesh_base_color, 1.0))
+
+        if cam_t is not None:
+            camera_translation = cam_t.copy()
+            # camera_translation[0] *= -1.
+        else:
+            camera_translation = np.array([0, 0, camera_z * self.focal_length/render_res[1]])
+
+        mesh = self.vertices_to_trimesh(vertices, camera_translation, mesh_base_color, rot_axis, rot_angle)
+        mesh = pyrender.Mesh.from_trimesh(mesh)
+        # mesh = pyrender.Mesh.from_trimesh(mesh, material=material)
+
+        scene = pyrender.Scene(bg_color=[*scene_bg_color, 0.0],
+                               ambient_light=(0.3, 0.3, 0.3))
+        scene.add(mesh, 'mesh')
+
+        camera_pose = np.eye(4)
+        # camera_pose[:3, 3] = camera_translation
+        camera_center = [render_res[0] / 2., render_res[1] / 2.]
+        camera = pyrender.IntrinsicsCamera(fx=self.focal_length, fy=self.focal_length,
+                                           cx=camera_center[0], cy=camera_center[1])
+
+        # Create camera node and add it to pyRender scene
+        camera_node = pyrender.Node(camera=camera, matrix=camera_pose)
+        scene.add_node(camera_node)
+        self.add_point_lighting(scene, camera_node)
+        self.add_lighting(scene, camera_node)
+
+        light_nodes = create_raymond_lights()
+        for node in light_nodes:
+            scene.add_node(node)
+
+        color, rend_depth = renderer.render(scene, flags=pyrender.RenderFlags.RGBA)
+        color = color.astype(np.float32) / 255.0
+        renderer.delete()
+
+        return color
+
+    def render_rgba_multiple(
+            self,
+            vertices: List[np.array],
+            cam_t: List[np.array],
+            rot_axis=[1,0,0],
+            rot_angle=0,
+            mesh_base_color=(1.0, 1.0, 0.9),
+            scene_bg_color=(0,0,0),
+            render_res=[256, 256],
+        ):
+
+        renderer = pyrender.OffscreenRenderer(viewport_width=render_res[0],
+                                              viewport_height=render_res[1],
+                                              point_size=1.0)
+        # material = pyrender.MetallicRoughnessMaterial(
+        #     metallicFactor=0.0,
+        #     alphaMode='OPAQUE',
+        #     baseColorFactor=(*mesh_base_color, 1.0))
+
+        mesh_list = [pyrender.Mesh.from_trimesh(self.vertices_to_trimesh(vvv, ttt.copy(), mesh_base_color, rot_axis, rot_angle)) for vvv,ttt in zip(vertices, cam_t)]
+
+        scene = pyrender.Scene(bg_color=[*scene_bg_color, 0.0],
+                               ambient_light=(0.3, 0.3, 0.3))
+        for i,mesh in enumerate(mesh_list):
+            scene.add(mesh, f'mesh_{i}')
+
+        camera_pose = np.eye(4)
+        # camera_pose[:3, 3] = camera_translation
+        camera_center = [render_res[0] / 2., render_res[1] / 2.]
+        camera = pyrender.IntrinsicsCamera(fx=self.focal_length, fy=self.focal_length,
+                                           cx=camera_center[0], cy=camera_center[1])
+
+        # Create camera node and add it to pyRender scene
+        camera_node = pyrender.Node(camera=camera, matrix=camera_pose)
+        scene.add_node(camera_node)
+        self.add_point_lighting(scene, camera_node)
+        self.add_lighting(scene, camera_node)
+
+        light_nodes = create_raymond_lights()
+        for node in light_nodes:
+            scene.add_node(node)
+
+        color, rend_depth = renderer.render(scene, flags=pyrender.RenderFlags.RGBA)
+        color = color.astype(np.float32) / 255.0
+        renderer.delete()
+
+        return color
+
+    def add_lighting(self, scene, cam_node, color=np.ones(3), intensity=1.0):
+        # from phalp.visualize.py_renderer import get_light_poses
+        light_poses = get_light_poses()
+        light_poses.append(np.eye(4))
+        cam_pose = scene.get_pose(cam_node)
+        for i, pose in enumerate(light_poses):
+            matrix = cam_pose @ pose
+            node = pyrender.Node(
+                name=f"light-{i:02d}",
+                light=pyrender.DirectionalLight(color=color, intensity=intensity),
+                matrix=matrix,
+            )
+            if scene.has_node(node):
+                continue
+            scene.add_node(node)
+
+    def add_point_lighting(self, scene, cam_node, color=np.ones(3), intensity=1.0):
+        # from phalp.visualize.py_renderer import get_light_poses
+        light_poses = get_light_poses(dist=0.5)
+        light_poses.append(np.eye(4))
+        cam_pose = scene.get_pose(cam_node)
+        for i, pose in enumerate(light_poses):
+            matrix = cam_pose @ pose
+            # node = pyrender.Node(
+            #     name=f"light-{i:02d}",
+            #     light=pyrender.DirectionalLight(color=color, intensity=intensity),
+            #     matrix=matrix,
+            # )
+            node = pyrender.Node(
+                name=f"plight-{i:02d}",
+                light=pyrender.PointLight(color=color, intensity=intensity),
+                matrix=matrix,
+            )
+            if scene.has_node(node):
+                continue
+            scene.add_node(node)

hmr2/utils/skeleton_renderer.py

@@ -0,0 +1,122 @@
+import torch
+import numpy as np
+import trimesh
+from typing import Optional
+from yacs.config import CfgNode
+
+from .geometry import perspective_projection
+from .render_openpose import render_openpose
+
+class SkeletonRenderer:
+
+    def __init__(self, cfg: CfgNode):
+        """
+        Object used to render 3D keypoints. Faster for use during training.
+        Args:
+            cfg (CfgNode): Model config file.
+        """
+        self.cfg = cfg
+
+    def __call__(self,
+                 pred_keypoints_3d: torch.Tensor,
+                 gt_keypoints_3d: torch.Tensor,
+                 gt_keypoints_2d: torch.Tensor,
+                 images: Optional[np.array] = None,
+                 camera_translation: Optional[torch.Tensor] = None) -> np.array:
+        """
+        Render batch of 3D keypoints.
+        Args:
+            pred_keypoints_3d (torch.Tensor): Tensor of shape (B, S, N, 3) containing a batch of predicted 3D keypoints, with S samples per image.
+            gt_keypoints_3d (torch.Tensor): Tensor of shape (B, N, 4) containing corresponding ground truth 3D keypoints; last value is the confidence.
+            gt_keypoints_2d (torch.Tensor): Tensor of shape (B, N, 3) containing corresponding ground truth 2D keypoints.
+            images (torch.Tensor): Tensor of shape (B, H, W, 3) containing images with values in the [0,255] range.
+            camera_translation (torch.Tensor): Tensor of shape (B, 3) containing the camera translation.
+        Returns:
+            np.array : Image with the following layout. Each row contains the a) input image,
+                                                                              b) image with gt 2D keypoints,
+                                                                              c) image with projected gt 3D keypoints,
+                                                                              d_1, ... , d_S) image with projected predicted 3D keypoints,
+                                                                              e) gt 3D keypoints rendered from a side view,
+                                                                              f_1, ... , f_S) predicted 3D keypoints frorm a side view
+        """
+        batch_size = pred_keypoints_3d.shape[0]
+#        num_samples = pred_keypoints_3d.shape[1]
+        pred_keypoints_3d = pred_keypoints_3d.clone().cpu().float()
+        gt_keypoints_3d = gt_keypoints_3d.clone().cpu().float()
+        gt_keypoints_3d[:, :, :-1] = gt_keypoints_3d[:, :, :-1] - gt_keypoints_3d[:, [25+14], :-1] + pred_keypoints_3d[:, [25+14]]
+        gt_keypoints_2d = gt_keypoints_2d.clone().cpu().float().numpy()
+        gt_keypoints_2d[:, :, :-1] = self.cfg.MODEL.IMAGE_SIZE * (gt_keypoints_2d[:, :, :-1] + 1.0) / 2.0
+
+        openpose_indices = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14]
+        gt_indices = [12, 8, 7, 6, 9, 10, 11, 14, 2, 1, 0, 3, 4, 5]
+        gt_indices = [25 + i for i in gt_indices]
+        keypoints_to_render = torch.ones(batch_size, gt_keypoints_3d.shape[1], 1)
+        rotation = torch.eye(3).unsqueeze(0)
+        if camera_translation is None:
+            camera_translation = torch.tensor([0.0, 0.0, 2 * self.cfg.EXTRA.FOCAL_LENGTH / (0.8 * self.cfg.MODEL.IMAGE_SIZE)]).unsqueeze(0).repeat(batch_size, 1)
+        else:
+            camera_translation = camera_translation.cpu()
+
+        if images is None:
+            images = np.zeros((batch_size, self.cfg.MODEL.IMAGE_SIZE, self.cfg.MODEL.IMAGE_SIZE, 3))
+        focal_length = torch.tensor([self.cfg.EXTRA.FOCAL_LENGTH, self.cfg.EXTRA.FOCAL_LENGTH]).reshape(1, 2)
+        camera_center = torch.tensor([self.cfg.MODEL.IMAGE_SIZE, self.cfg.MODEL.IMAGE_SIZE], dtype=torch.float).reshape(1, 2) / 2.
+        gt_keypoints_3d_proj = perspective_projection(gt_keypoints_3d[:, :, :-1], rotation=rotation.repeat(batch_size, 1, 1), translation=camera_translation[:, :], focal_length=focal_length.repeat(batch_size, 1), camera_center=camera_center.repeat(batch_size, 1))
+        pred_keypoints_3d_proj = perspective_projection(pred_keypoints_3d.reshape(batch_size, -1, 3), rotation=rotation.repeat(batch_size, 1, 1), translation=camera_translation.reshape(batch_size, -1), focal_length=focal_length.repeat(batch_size, 1), camera_center=camera_center.repeat(batch_size, 1)).reshape(batch_size, -1, 2)
+        gt_keypoints_3d_proj = torch.cat([gt_keypoints_3d_proj, gt_keypoints_3d[:, :, [-1]]], dim=-1).cpu().numpy()
+        pred_keypoints_3d_proj = torch.cat([pred_keypoints_3d_proj, keypoints_to_render.reshape(batch_size, -1, 1)], dim=-1).cpu().numpy()
+        rows = []
+        # Rotate keypoints to visualize side view
+        R = torch.tensor(trimesh.transformations.rotation_matrix(np.radians(90), [0, 1, 0])[:3, :3]).float()
+        gt_keypoints_3d_side = gt_keypoints_3d.clone()
+        gt_keypoints_3d_side[:, :, :-1] = torch.einsum('bni,ij->bnj', gt_keypoints_3d_side[:, :, :-1], R)
+        pred_keypoints_3d_side = pred_keypoints_3d.clone()
+        pred_keypoints_3d_side = torch.einsum('bni,ij->bnj', pred_keypoints_3d_side, R)
+        gt_keypoints_3d_proj_side = perspective_projection(gt_keypoints_3d_side[:, :, :-1], rotation=rotation.repeat(batch_size, 1, 1), translation=camera_translation[:, :], focal_length=focal_length.repeat(batch_size, 1), camera_center=camera_center.repeat(batch_size, 1))
+        pred_keypoints_3d_proj_side = perspective_projection(pred_keypoints_3d_side.reshape(batch_size, -1, 3), rotation=rotation.repeat(batch_size, 1, 1), translation=camera_translation.reshape(batch_size, -1), focal_length=focal_length.repeat(batch_size, 1), camera_center=camera_center.repeat(batch_size, 1)).reshape(batch_size, -1, 2)
+        gt_keypoints_3d_proj_side = torch.cat([gt_keypoints_3d_proj_side, gt_keypoints_3d_side[:, :, [-1]]], dim=-1).cpu().numpy()
+        pred_keypoints_3d_proj_side = torch.cat([pred_keypoints_3d_proj_side, keypoints_to_render.reshape(batch_size, -1, 1)], dim=-1).cpu().numpy()
+        for i in range(batch_size):
+            img = images[i]
+            side_img = np.zeros((self.cfg.MODEL.IMAGE_SIZE, self.cfg.MODEL.IMAGE_SIZE, 3))
+            # gt 2D keypoints
+            body_keypoints_2d = gt_keypoints_2d[i, :25].copy()
+            for op, gt in zip(openpose_indices, gt_indices):
+                if gt_keypoints_2d[i, gt, -1] > body_keypoints_2d[op, -1]:
+                    body_keypoints_2d[op] = gt_keypoints_2d[i, gt]
+            gt_keypoints_img = render_openpose(img, body_keypoints_2d) / 255.
+            # gt 3D keypoints
+            body_keypoints_3d_proj = gt_keypoints_3d_proj[i, :25].copy()
+            for op, gt in zip(openpose_indices, gt_indices):
+                if gt_keypoints_3d_proj[i, gt, -1] > body_keypoints_3d_proj[op, -1]:
+                    body_keypoints_3d_proj[op] = gt_keypoints_3d_proj[i, gt]
+            gt_keypoints_3d_proj_img = render_openpose(img, body_keypoints_3d_proj) / 255.
+            # gt 3D keypoints from the side
+            body_keypoints_3d_proj = gt_keypoints_3d_proj_side[i, :25].copy()
+            for op, gt in zip(openpose_indices, gt_indices):
+                if gt_keypoints_3d_proj_side[i, gt, -1] > body_keypoints_3d_proj[op, -1]:
+                    body_keypoints_3d_proj[op] = gt_keypoints_3d_proj_side[i, gt]
+            gt_keypoints_3d_proj_img_side = render_openpose(side_img, body_keypoints_3d_proj) / 255.
+            # pred 3D keypoints
+            pred_keypoints_3d_proj_imgs = []
+            body_keypoints_3d_proj = pred_keypoints_3d_proj[i, :25].copy()
+            for op, gt in zip(openpose_indices, gt_indices):
+                if pred_keypoints_3d_proj[i, gt, -1] >= body_keypoints_3d_proj[op, -1]:
+                    body_keypoints_3d_proj[op] = pred_keypoints_3d_proj[i, gt]
+            pred_keypoints_3d_proj_imgs.append(render_openpose(img, body_keypoints_3d_proj) / 255.)
+            pred_keypoints_3d_proj_img = np.concatenate(pred_keypoints_3d_proj_imgs, axis=1)
+            # gt 3D keypoints from the side
+            pred_keypoints_3d_proj_imgs_side = []
+            body_keypoints_3d_proj = pred_keypoints_3d_proj_side[i, :25].copy()
+            for op, gt in zip(openpose_indices, gt_indices):
+                if pred_keypoints_3d_proj_side[i, gt, -1] >= body_keypoints_3d_proj[op, -1]:
+                    body_keypoints_3d_proj[op] = pred_keypoints_3d_proj_side[i, gt]
+            pred_keypoints_3d_proj_imgs_side.append(render_openpose(side_img, body_keypoints_3d_proj) / 255.)
+            pred_keypoints_3d_proj_img_side = np.concatenate(pred_keypoints_3d_proj_imgs_side, axis=1)
+            rows.append(np.concatenate((gt_keypoints_img, gt_keypoints_3d_proj_img, pred_keypoints_3d_proj_img, gt_keypoints_3d_proj_img_side, pred_keypoints_3d_proj_img_side), axis=1))
+        # Concatenate images
+        img = np.concatenate(rows, axis=0)
+        img[:, ::self.cfg.MODEL.IMAGE_SIZE, :] = 1.0
+        img[::self.cfg.MODEL.IMAGE_SIZE, :, :] = 1.0
+        img[:, (1+1+1)*self.cfg.MODEL.IMAGE_SIZE, :] = 0.5
+        return img

hmr2/utils/texture_utils.py

@@ -0,0 +1,85 @@
+import numpy as np
+import torch
+from torch.nn import functional as F
+# from psbody.mesh.visibility import visibility_compute
+
+def uv_to_xyz_and_normals(verts, f, fmap, bmap, ftov):
+    vn = estimate_vertex_normals(verts, f, ftov)
+    pixels_to_set = torch.nonzero(fmap+1)
+    x_to_set = pixels_to_set[:,0]
+    y_to_set = pixels_to_set[:,1]
+    b_coords = bmap[x_to_set, y_to_set, :]
+    f_coords = fmap[x_to_set, y_to_set]
+    v_ids = f[f_coords]
+    points = (b_coords[:,0,None]*verts[:,v_ids[:,0]]
+             + b_coords[:,1,None]*verts[:,v_ids[:,1]]
+             + b_coords[:,2,None]*verts[:,v_ids[:,2]])
+    normals = (b_coords[:,0,None]*vn[:,v_ids[:,0]]
+             + b_coords[:,1,None]*vn[:,v_ids[:,1]]
+             + b_coords[:,2,None]*vn[:,v_ids[:,2]])
+    return points, normals, vn, f_coords
+
+def estimate_vertex_normals(v, f, ftov):
+    face_normals = TriNormalsScaled(v, f)
+    non_scaled_normals = torch.einsum('ij,bjk->bik', ftov, face_normals)
+    norms = torch.sum(non_scaled_normals ** 2.0, 2) ** 0.5
+    norms[norms == 0] = 1.0
+    return torch.div(non_scaled_normals, norms[:,:,None])
+
+def TriNormalsScaled(v, f):
+    return torch.cross(_edges_for(v, f, 1, 0), _edges_for(v, f, 2, 0))
+
+def _edges_for(v, f, cplus, cminus):
+    return v[:,f[:,cplus]] - v[:,f[:,cminus]]
+
+def psbody_get_face_visibility(v, n, f, cams, normal_threshold=0.5):
+    bn, nverts, _ = v.shape
+    nfaces, _ = f.shape
+    vis_f = np.zeros([bn, nfaces], dtype='float32')
+    for i in range(bn):
+        vis, n_dot_cam = visibility_compute(v=v[i], n=n[i], f=f, cams=cams)
+        vis_v = (vis == 1) & (n_dot_cam > normal_threshold)
+        vis_f[i] = np.all(vis_v[0,f],1)
+    return vis_f
+
+def compute_uvsampler(vt, ft, tex_size=6):
+    """
+    For this mesh, pre-computes the UV coordinates for
+    F x T x T points.
+    Returns F x T x T x 2
+    """
+    uv = obj2nmr_uvmap(ft, vt, tex_size=tex_size)
+    uv = uv.reshape(-1, tex_size, tex_size, 2)
+    return uv
+
+def obj2nmr_uvmap(ft, vt, tex_size=6):
+    """
+    Converts obj uv_map to NMR uv_map (F x T x T x 2),
+    where tex_size (T) is the sample rate on each face.
+    """
+    # This is F x 3 x 2
+    uv_map_for_verts = vt[ft]
+
+    # obj's y coordinate is [1-0], but image is [0-1]
+    uv_map_for_verts[:, :, 1] = 1 - uv_map_for_verts[:, :, 1]
+
+    # range [0, 1] -> [-1, 1]
+    uv_map_for_verts = (2 * uv_map_for_verts) - 1
+
+    alpha = np.arange(tex_size, dtype=np.float) / (tex_size - 1)
+    beta = np.arange(tex_size, dtype=np.float) / (tex_size - 1)
+    import itertools
+    # Barycentric coordinate values
+    coords = np.stack([p for p in itertools.product(*[alpha, beta])])
+
+    # Compute alpha, beta (this is the same order as NMR)
+    v2 = uv_map_for_verts[:, 2]
+    v0v2 = uv_map_for_verts[:, 0] - uv_map_for_verts[:, 2]
+    v1v2 = uv_map_for_verts[:, 1] - uv_map_for_verts[:, 2]
+    # Interpolate the vertex uv values: F x 2 x T*2
+    uv_map = np.dstack([v0v2, v1v2]).dot(coords.T) + v2.reshape(-1, 2, 1)
+
+    # F x T*2 x 2  -> F x T x T x 2
+    uv_map = np.transpose(uv_map, (0, 2, 1)).reshape(-1, tex_size, tex_size, 2)
+
+    return uv_map

hmr2/utils/utils_detectron2.py

@@ -0,0 +1,93 @@
+import detectron2.data.transforms as T
+import torch
+from detectron2.checkpoint import DetectionCheckpointer
+from detectron2.config import CfgNode, instantiate
+from detectron2.data import MetadataCatalog
+from omegaconf import OmegaConf
+
+
+class DefaultPredictor_Lazy:
+    """Create a simple end-to-end predictor with the given config that runs on single device for a
+    single input image.
+
+    Compared to using the model directly, this class does the following additions:
+
+    1. Load checkpoint from the weights specified in config (cfg.MODEL.WEIGHTS).
+    2. Always take BGR image as the input and apply format conversion internally.
+    3. Apply resizing defined by the config (`cfg.INPUT.{MIN,MAX}_SIZE_TEST`).
+    4. Take one input image and produce a single output, instead of a batch.
+
+    This is meant for simple demo purposes, so it does the above steps automatically.
+    This is not meant for benchmarks or running complicated inference logic.
+    If you'd like to do anything more complicated, please refer to its source code as
+    examples to build and use the model manually.
+
+    Attributes:
+        metadata (Metadata): the metadata of the underlying dataset, obtained from
+            test dataset name in the config.
+
+
+    Examples:
+    ::
+        pred = DefaultPredictor(cfg)
+        inputs = cv2.imread("input.jpg")
+        outputs = pred(inputs)
+    """
+
+    def __init__(self, cfg):
+        """
+        Args:
+            cfg: a yacs CfgNode or a omegaconf dict object.
+        """
+        if isinstance(cfg, CfgNode):
+            self.cfg = cfg.clone()  # cfg can be modified by model
+            self.model = build_model(self.cfg)  # noqa: F821
+            if len(cfg.DATASETS.TEST):
+                test_dataset = cfg.DATASETS.TEST[0]
+
+            checkpointer = DetectionCheckpointer(self.model)
+            checkpointer.load(cfg.MODEL.WEIGHTS)
+
+            self.aug = T.ResizeShortestEdge(
+                [cfg.INPUT.MIN_SIZE_TEST, cfg.INPUT.MIN_SIZE_TEST], cfg.INPUT.MAX_SIZE_TEST
+            )
+
+            self.input_format = cfg.INPUT.FORMAT
+        else:  # new LazyConfig
+            self.cfg = cfg
+            self.model = instantiate(cfg.model)
+            test_dataset = OmegaConf.select(cfg, "dataloader.test.dataset.names", default=None)
+            if isinstance(test_dataset, (list, tuple)):
+                test_dataset = test_dataset[0]
+
+            checkpointer = DetectionCheckpointer(self.model)
+            checkpointer.load(OmegaConf.select(cfg, "train.init_checkpoint", default=""))
+
+            mapper = instantiate(cfg.dataloader.test.mapper)
+            self.aug = mapper.augmentations
+            self.input_format = mapper.image_format
+
+        self.model.eval().cuda()
+        if test_dataset:
+            self.metadata = MetadataCatalog.get(test_dataset)
+        assert self.input_format in ["RGB", "BGR"], self.input_format
+
+    def __call__(self, original_image):
+        """
+        Args:
+            original_image (np.ndarray): an image of shape (H, W, C) (in BGR order).
+
+        Returns:
+            predictions (dict):
+                the output of the model for one image only.
+                See :doc:`/tutorials/models` for details about the format.
+        """
+        with torch.no_grad():
+            if self.input_format == "RGB":
+                original_image = original_image[:, :, ::-1]
+            height, width = original_image.shape[:2]
+            image = self.aug(T.AugInput(original_image)).apply_image(original_image)
+            image = torch.as_tensor(image.astype("float32").transpose(2, 0, 1))
+            inputs = {"image": image, "height": height, "width": width}
+            predictions = self.model([inputs])[0]
+            return predictions

requirements.txt

@@ -0,0 +1,29 @@
+# torch
+# pytorch-lightning
+# smplx==0.1.28
+# pyrender
+# opencv-python
+# yacs
+# scikit-image
+# einops
+# timm
+# OmegaConf
+
+--extra-index-url https://download.pytorch.org/whl/cu116
+torch==1.13.1+cu116
+torchvision==0.14.1+cu116
+pytorch-lightning
+smplx==0.1.28
+opencv-python
+yacs
+scikit-image
+einops
+timm
+OmegaConf
+trimesh
+pyopengl==3.1.0
+pyglet
+PyOpenGL
+PyOpenGL_accelerate
+numpy==1.23.3
+shapely 

setup.py

@@ -0,0 +1,8 @@
+from setuptools import setup, find_packages
+
+print('Found packages:', find_packages())
+setup(
+    description='HMR2 as a package',
+    name='hmr2',
+    packages=find_packages()
+)

vendor/detectron2/.circleci/config.yml

@@ -0,0 +1,271 @@
+version: 2.1
+
+# -------------------------------------------------------------------------------------
+# Environments to run the jobs in
+# -------------------------------------------------------------------------------------
+cpu: &cpu
+  machine:
+    image: ubuntu-2004:202107-02
+  resource_class: medium
+
+gpu: &gpu
+  machine:
+    # NOTE: use a cuda version that's supported by all our pytorch versions
+    image: ubuntu-1604-cuda-11.1:202012-01
+  resource_class: gpu.nvidia.small
+
+windows-cpu: &windows_cpu
+  machine:
+    resource_class: windows.medium
+    image: windows-server-2019-vs2019:stable
+    shell: powershell.exe
+
+# windows-gpu: &windows_gpu
+#     machine:
+#       resource_class: windows.gpu.nvidia.medium
+#       image: windows-server-2019-nvidia:stable
+
+version_parameters: &version_parameters
+  parameters:
+    pytorch_version:
+      type: string
+    torchvision_version:
+      type: string
+    pytorch_index:
+      type: string
+      # use test wheels index to have access to RC wheels
+      # https://download.pytorch.org/whl/test/torch_test.html
+      default: "https://download.pytorch.org/whl/torch_stable.html"
+    python_version:  # NOTE: only affect linux
+      type: string
+      default: '3.8.6'
+
+  environment:
+    PYTORCH_VERSION: << parameters.pytorch_version >>
+    TORCHVISION_VERSION: << parameters.torchvision_version >>
+    PYTORCH_INDEX: << parameters.pytorch_index >>
+    PYTHON_VERSION: << parameters.python_version>>
+    # point datasets to ~/.torch so it's cached in CI
+    DETECTRON2_DATASETS: ~/.torch/datasets
+
+# -------------------------------------------------------------------------------------
+# Re-usable commands
+# -------------------------------------------------------------------------------------
+# install_nvidia_driver: &install_nvidia_driver
+#   - run:
+#       name: Install nvidia driver
+#       working_directory: ~/
+#       command: |
+#         wget -q 'https://s3.amazonaws.com/ossci-linux/nvidia_driver/NVIDIA-Linux-x86_64-430.40.run'
+#         sudo /bin/bash ./NVIDIA-Linux-x86_64-430.40.run -s --no-drm
+#         nvidia-smi
+
+add_ssh_keys: &add_ssh_keys
+  # https://circleci.com/docs/2.0/add-ssh-key/
+  - add_ssh_keys:
+      fingerprints:
+        - "e4:13:f2:22:d4:49:e8:e4:57:5a:ac:20:2f:3f:1f:ca"
+
+install_python: &install_python
+  - run:
+      name: Install Python
+      working_directory: ~/
+      command: |
+        # upgrade pyenv
+        cd /opt/circleci/.pyenv/plugins/python-build/../.. && git pull && cd -
+        pyenv install -s $PYTHON_VERSION
+        pyenv global $PYTHON_VERSION
+        python --version
+        which python
+        pip install --upgrade pip
+
+setup_venv: &setup_venv
+  - run:
+      name: Setup Virtual Env
+      working_directory: ~/
+      command: |
+        python -m venv ~/venv
+        echo ". ~/venv/bin/activate" >> $BASH_ENV
+        . ~/venv/bin/activate
+        python --version
+        which python
+        which pip
+        pip install --upgrade pip
+
+setup_venv_win: &setup_venv_win
+  - run:
+      name: Setup Virtual Env for Windows
+      command: |
+        pip install virtualenv
+        python -m virtualenv env
+        .\env\Scripts\activate
+        python --version
+        which python
+        which pip
+
+install_linux_dep: &install_linux_dep
+  - run:
+      name: Install Dependencies
+      command: |
+        # disable crash coredump, so unittests fail fast
+        sudo systemctl stop apport.service
+        # install from github to get latest; install iopath first since fvcore depends on it
+        pip install --progress-bar off -U 'git+https://github.com/facebookresearch/iopath'
+        pip install --progress-bar off -U 'git+https://github.com/facebookresearch/fvcore'
+        # Don't use pytest-xdist: cuda tests are unstable under multi-process workers.
+        # Don't use opencv 4.7.0.68: https://github.com/opencv/opencv-python/issues/765
+        pip install --progress-bar off ninja opencv-python-headless!=4.7.0.68 pytest tensorboard pycocotools onnx
+        pip install --progress-bar off torch==$PYTORCH_VERSION -f $PYTORCH_INDEX
+        if [[ "$TORCHVISION_VERSION" == "master" ]]; then
+          pip install git+https://github.com/pytorch/vision.git
+        else
+          pip install --progress-bar off torchvision==$TORCHVISION_VERSION -f $PYTORCH_INDEX
+        fi
+
+        python -c 'import torch; print("CUDA:", torch.cuda.is_available())'
+        gcc --version
+
+install_detectron2: &install_detectron2
+  - run:
+      name: Install Detectron2
+      command: |
+        # Remove first, in case it's in the CI cache
+        pip uninstall -y detectron2
+
+        pip install --progress-bar off -e .[all]
+        python -m detectron2.utils.collect_env
+        ./datasets/prepare_for_tests.sh
+
+run_unittests: &run_unittests
+  - run:
+      name: Run Unit Tests
+      command: |
+        pytest -sv --durations=15 tests  # parallel causes some random failures
+
+uninstall_tests: &uninstall_tests
+  - run:
+      name: Run Tests After Uninstalling
+      command: |
+        pip uninstall -y detectron2
+        # Remove built binaries
+        rm -rf build/ detectron2/*.so
+        # Tests that code is importable without installation
+        PYTHONPATH=. ./.circleci/import-tests.sh
+
+
+# -------------------------------------------------------------------------------------
+# Jobs to run
+# -------------------------------------------------------------------------------------
+jobs:
+  linux_cpu_tests:
+    <<: *cpu
+    <<: *version_parameters
+
+    working_directory: ~/detectron2
+
+    steps:
+      - checkout
+
+      # Cache the venv directory that contains python, dependencies, and checkpoints
+      # Refresh the key when dependencies should be updated (e.g. when pytorch releases)
+      - restore_cache:
+          keys:
+            - cache-{{ arch }}-<< parameters.pytorch_version >>-{{ .Branch }}-20210827
+
+      - <<: *install_python
+      - <<: *install_linux_dep
+      - <<: *install_detectron2
+      - <<: *run_unittests
+      - <<: *uninstall_tests
+
+      - save_cache:
+          paths:
+            - /opt/circleci/.pyenv
+            - ~/.torch
+          key: cache-{{ arch }}-<< parameters.pytorch_version >>-{{ .Branch }}-20210827
+
+
+  linux_gpu_tests:
+    <<: *gpu
+    <<: *version_parameters
+
+    working_directory: ~/detectron2
+
+    steps:
+      - checkout
+
+      - restore_cache:
+          keys:
+            - cache-{{ arch }}-<< parameters.pytorch_version >>-{{ .Branch }}-20210827
+
+      - <<: *install_python
+      - <<: *install_linux_dep
+      - <<: *install_detectron2
+      - <<: *run_unittests
+      - <<: *uninstall_tests
+
+      - save_cache:
+          paths:
+            - /opt/circleci/.pyenv
+            - ~/.torch
+          key: cache-{{ arch }}-<< parameters.pytorch_version >>-{{ .Branch }}-20210827
+
+  windows_cpu_build:
+    <<: *windows_cpu
+    <<: *version_parameters
+    steps:
+      - <<: *add_ssh_keys
+      - checkout
+      - <<: *setup_venv_win
+
+      # Cache the env directory that contains dependencies
+      - restore_cache:
+          keys:
+            - cache-{{ arch }}-<< parameters.pytorch_version >>-{{ .Branch }}-20210404
+
+      - run:
+          name: Install Dependencies
+          command: |
+            pip install certifi --ignore-installed  # required on windows to workaround some cert issue
+            pip install numpy cython  # required on windows before pycocotools
+            pip install opencv-python-headless pytest-xdist pycocotools tensorboard onnx
+            pip install -U git+https://github.com/facebookresearch/iopath
+            pip install -U git+https://github.com/facebookresearch/fvcore
+            pip install torch==$env:PYTORCH_VERSION torchvision==$env:TORCHVISION_VERSION -f $env:PYTORCH_INDEX
+
+      - save_cache:
+          paths:
+            - env
+          key: cache-{{ arch }}-<< parameters.pytorch_version >>-{{ .Branch }}-20210404
+
+      - <<: *install_detectron2
+      # TODO: unittest fails for now
+
+workflows:
+  version: 2
+  regular_test:
+    jobs:
+      - linux_cpu_tests:
+          name: linux_cpu_tests_pytorch1.10
+          pytorch_version: '1.10.0+cpu'
+          torchvision_version: '0.11.1+cpu'
+      - linux_gpu_tests:
+          name: linux_gpu_tests_pytorch1.8
+          pytorch_version: '1.8.1+cu111'
+          torchvision_version: '0.9.1+cu111'
+      - linux_gpu_tests:
+          name: linux_gpu_tests_pytorch1.9
+          pytorch_version: '1.9+cu111'
+          torchvision_version: '0.10+cu111'
+      - linux_gpu_tests:
+          name: linux_gpu_tests_pytorch1.10
+          pytorch_version: '1.10+cu111'
+          torchvision_version: '0.11.1+cu111'
+      - linux_gpu_tests:
+          name: linux_gpu_tests_pytorch1.10_python39
+          pytorch_version: '1.10+cu111'
+          torchvision_version: '0.11.1+cu111'
+          python_version: '3.9.6'
+      - windows_cpu_build:
+          pytorch_version: '1.10+cpu'
+          torchvision_version: '0.11.1+cpu'

vendor/detectron2/.circleci/import-tests.sh

@@ -0,0 +1,16 @@
+#!/bin/bash -e
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+# Test that import works without building detectron2.
+
+# Check that _C is not importable
+python -c "from detectron2 import _C" > /dev/null 2>&1 && {
+  echo "This test should be run without building detectron2."
+  exit 1
+}
+
+# Check that other modules are still importable, even when _C is not importable
+python -c "from detectron2 import modeling"
+python -c "from detectron2 import modeling, data"
+python -c "from detectron2 import evaluation, export, checkpoint"
+python -c "from detectron2 import utils, engine"

vendor/detectron2/.clang-format

@@ -0,0 +1,85 @@
+AccessModifierOffset: -1
+AlignAfterOpenBracket: AlwaysBreak
+AlignConsecutiveAssignments: false
+AlignConsecutiveDeclarations: false
+AlignEscapedNewlinesLeft: true
+AlignOperands:   false
+AlignTrailingComments: false
+AllowAllParametersOfDeclarationOnNextLine: false
+AllowShortBlocksOnASingleLine: false
+AllowShortCaseLabelsOnASingleLine: false
+AllowShortFunctionsOnASingleLine: Empty
+AllowShortIfStatementsOnASingleLine: false
+AllowShortLoopsOnASingleLine: false
+AlwaysBreakAfterReturnType: None
+AlwaysBreakBeforeMultilineStrings: true
+AlwaysBreakTemplateDeclarations: true
+BinPackArguments: false
+BinPackParameters: false
+BraceWrapping:
+  AfterClass:      false
+  AfterControlStatement: false
+  AfterEnum:       false
+  AfterFunction:   false
+  AfterNamespace:  false
+  AfterObjCDeclaration: false
+  AfterStruct:     false
+  AfterUnion:      false
+  BeforeCatch:     false
+  BeforeElse:      false
+  IndentBraces:    false
+BreakBeforeBinaryOperators: None
+BreakBeforeBraces: Attach
+BreakBeforeTernaryOperators: true
+BreakConstructorInitializersBeforeComma: false
+BreakAfterJavaFieldAnnotations: false
+BreakStringLiterals: false
+ColumnLimit:     80
+CommentPragmas:  '^ IWYU pragma:'
+ConstructorInitializerAllOnOneLineOrOnePerLine: true
+ConstructorInitializerIndentWidth: 4
+ContinuationIndentWidth: 4
+Cpp11BracedListStyle: true
+DerivePointerAlignment: false
+DisableFormat:   false
+ForEachMacros:   [ FOR_EACH, FOR_EACH_R, FOR_EACH_RANGE, ]
+IncludeCategories:
+  - Regex:           '^<.*\.h(pp)?>'
+    Priority:        1
+  - Regex:           '^<.*'
+    Priority:        2
+  - Regex:           '.*'
+    Priority:        3
+IndentCaseLabels: true
+IndentWidth:     2
+IndentWrappedFunctionNames: false
+KeepEmptyLinesAtTheStartOfBlocks: false
+MacroBlockBegin: ''
+MacroBlockEnd:   ''
+MaxEmptyLinesToKeep: 1
+NamespaceIndentation: None
+ObjCBlockIndentWidth: 2
+ObjCSpaceAfterProperty: false
+ObjCSpaceBeforeProtocolList: false
+PenaltyBreakBeforeFirstCallParameter: 1
+PenaltyBreakComment: 300
+PenaltyBreakFirstLessLess: 120
+PenaltyBreakString: 1000
+PenaltyExcessCharacter: 1000000
+PenaltyReturnTypeOnItsOwnLine: 200
+PointerAlignment: Left
+ReflowComments:  true
+SortIncludes:    true
+SpaceAfterCStyleCast: false
+SpaceBeforeAssignmentOperators: true
+SpaceBeforeParens: ControlStatements
+SpaceInEmptyParentheses: false
+SpacesBeforeTrailingComments: 1
+SpacesInAngles:  false
+SpacesInContainerLiterals: true
+SpacesInCStyleCastParentheses: false
+SpacesInParentheses: false
+SpacesInSquareBrackets: false
+Standard:        Cpp11
+TabWidth:        8
+UseTab:          Never

vendor/detectron2/.flake8

@@ -0,0 +1,15 @@
+# This is an example .flake8 config, used when developing *Black* itself.
+# Keep in sync with setup.cfg which is used for source packages.
+
+[flake8]
+ignore = W503, E203, E221, C901, C408, E741, C407, B017, F811, C101, EXE001, EXE002
+max-line-length = 100
+max-complexity = 18
+select = B,C,E,F,W,T4,B9
+exclude = build
+per-file-ignores =
+  **/__init__.py:F401,F403,E402
+  **/configs/**.py:F401,E402
+  configs/**.py:F401,E402
+  **/tests/config/**.py:F401,E402
+  tests/config/**.py:F401,E402

vendor/detectron2/.github/CODE_OF_CONDUCT.md

@@ -0,0 +1,5 @@
+# Code of Conduct
+
+Facebook has adopted a Code of Conduct that we expect project participants to adhere to.
+Please read the [full text](https://code.fb.com/codeofconduct/)
+so that you can understand what actions will and will not be tolerated.

vendor/detectron2/.github/CONTRIBUTING.md

@@ -0,0 +1,68 @@
+# Contributing to detectron2
+
+## Issues
+We use GitHub issues to track public bugs and questions.
+Please make sure to follow one of the
+[issue templates](https://github.com/facebookresearch/detectron2/issues/new/choose)
+when reporting any issues.
+
+Facebook has a [bounty program](https://www.facebook.com/whitehat/) for the safe
+disclosure of security bugs. In those cases, please go through the process
+outlined on that page and do not file a public issue.
+
+## Pull Requests
+We actively welcome pull requests.
+
+However, if you're adding any significant features (e.g. > 50 lines), please
+make sure to discuss with maintainers about your motivation and proposals in an issue
+before sending a PR. This is to save your time so you don't spend time on a PR that we'll not accept.
+
+We do not always accept new features, and we take the following
+factors into consideration:
+
+1. Whether the same feature can be achieved without modifying detectron2.
+   Detectron2 is designed so that you can implement many extensions from the outside, e.g.
+   those in [projects](https://github.com/facebookresearch/detectron2/tree/master/projects).
+   * If some part of detectron2 is not extensible enough, you can also bring up a more general issue to
+     improve it. Such feature request may be useful to more users.
+2. Whether the feature is potentially useful to a large audience (e.g. an impactful detection paper, a popular dataset,
+   a significant speedup, a widely useful utility),
+   or only to a small portion of users (e.g., a less-known paper, an improvement not in the object
+   detection field, a trick that's not very popular in the community, code to handle a non-standard type of data)
+   * Adoption of additional models, datasets, new task are by default not added to detectron2 before they
+     receive significant popularity in the community.
+     We sometimes accept such features in `projects/`, or as a link in `projects/README.md`.
+3. Whether the proposed solution has a good design / interface. This can be discussed in the issue prior to PRs, or
+   in the form of a draft PR.
+4. Whether the proposed solution adds extra mental/practical overhead to users who don't
+   need such feature.
+5. Whether the proposed solution breaks existing APIs.
+
+To add a feature to an existing function/class `Func`, there are always two approaches:
+(1) add new arguments to `Func`; (2) write a new `Func_with_new_feature`.
+To meet the above criteria, we often prefer approach (2), because:
+
+1. It does not involve modifying or potentially breaking existing code.
+2. It does not add overhead to users who do not need the new feature.
+3. Adding new arguments to a function/class is not scalable w.r.t. all the possible new research ideas in the future.
+
+When sending a PR, please do:
+
+1. If a PR contains multiple orthogonal changes, split it to several PRs.
+2. If you've added code that should be tested, add tests.
+3. For PRs that need experiments (e.g. adding a new model or new methods),
+   you don't need to update model zoo, but do provide experiment results in the description of the PR.
+4. If APIs are changed, update the documentation.
+5. We use the [Google style docstrings](https://www.sphinx-doc.org/en/master/usage/extensions/napoleon.html) in python.
+6. Make sure your code lints with `./dev/linter.sh`.
+
+
+## Contributor License Agreement ("CLA")
+In order to accept your pull request, we need you to submit a CLA. You only need
+to do this once to work on any of Facebook's open source projects.
+
+Complete your CLA here: <https://code.facebook.com/cla>
+
+## License
+By contributing to detectron2, you agree that your contributions will be licensed
+under the LICENSE file in the root directory of this source tree.

vendor/detectron2/.github/ISSUE_TEMPLATE.md

@@ -0,0 +1,5 @@
+
+Please select an issue template from
+https://github.com/facebookresearch/detectron2/issues/new/choose .
+
+Otherwise your issue will be closed.

vendor/detectron2/.github/ISSUE_TEMPLATE/bugs.md

@@ -0,0 +1,38 @@
+---
+name: "🐛 Bugs"
+about: Report bugs in detectron2
+title: Please read & provide the following
+
+---
+
+## Instructions To Reproduce the 🐛 Bug:
+1. Full runnable code or full changes you made:
+```
+If making changes to the project itself, please use output of the following command:
+git rev-parse HEAD; git diff
+
+<put code or diff here>
+```
+2. What exact command you run:
+3. __Full logs__ or other relevant observations:
+```
+<put logs here>
+```
+4. please simplify the steps as much as possible so they do not require additional resources to
+   run, such as a private dataset.
+
+## Expected behavior:
+
+If there are no obvious error in "full logs" provided above,
+please tell us the expected behavior.
+
+## Environment:
+
+Provide your environment information using the following command:
+```
+wget -nc -q https://github.com/facebookresearch/detectron2/raw/main/detectron2/utils/collect_env.py && python collect_env.py
+```
+
+If your issue looks like an installation issue / environment issue,
+please first try to solve it yourself with the instructions in
+https://detectron2.readthedocs.io/tutorials/install.html#common-installation-issues

vendor/detectron2/.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,17 @@
+# require an issue template to be chosen
+blank_issues_enabled: false
+
+contact_links:
+  - name: How-To / All Other Questions
+    url: https://github.com/facebookresearch/detectron2/discussions
+    about: Use "github discussions" for community support on general questions that don't belong to the above issue categories
+  - name: Detectron2 Documentation
+    url: https://detectron2.readthedocs.io/index.html
+    about: Check if your question is answered in tutorials or API docs
+
+# Unexpected behaviors & bugs are split to two templates.
+# When they are one template, users think "it's not a bug" and don't choose the template.
+#
+# But the file name is still "unexpected-problems-bugs.md" so that old references
+# to this issue template still works.
+# It's ok since this template should be a superset of "bugs.md" (unexpected behaviors is a superset of bugs)

vendor/detectron2/.github/ISSUE_TEMPLATE/documentation.md

@@ -0,0 +1,14 @@
+---
+name: "\U0001F4DA Documentation Issue"
+about: Report a problem about existing documentation, comments, website or tutorials.
+labels: documentation
+
+---
+
+## 📚 Documentation Issue
+
+This issue category is for problems about existing documentation, not for asking how-to questions.
+
+* Provide a link to an existing documentation/comment/tutorial:
+
+* How should the above documentation/comment/tutorial improve:

vendor/detectron2/.github/ISSUE_TEMPLATE/feature-request.md

@@ -0,0 +1,31 @@
+---
+name: "\U0001F680Feature Request"
+about: Suggest an improvement or new feature
+labels: enhancement
+
+---
+
+## 🚀 Feature
+A clear and concise description of the feature proposal.
+
+## Motivation & Examples
+
+Tell us why the feature is useful.
+
+Describe what the feature would look like, if it is implemented.
+Best demonstrated using **code examples** in addition to words.
+
+## Note
+
+We only consider adding new features if they are relevant to many users.
+
+If you request implementation of research papers -- we only consider papers that have enough significance and prevalance in the object detection field.
+
+We do not take requests for most projects in the `projects/` directory, because they are research code release that is mainly for other researchers to reproduce results.
+
+"Make X faster/accurate" is not a valid feature request. "Implement a concrete feature that can make X faster/accurate" can be a valid feature request.
+
+Instead of adding features inside detectron2,
+you can implement many features by [extending detectron2](https://detectron2.readthedocs.io/tutorials/extend.html).
+The [projects/](https://github.com/facebookresearch/detectron2/tree/main/projects/) directory contains many of such examples.
+