gradio_demo/barc_demo_v3.py

# python gradio_demo/barc_demo_v3.py

import numpy as np
import os
import glob
import torch
from torch.utils.data import DataLoader
import torchvision
from torchvision.models.detection.faster_rcnn import FastRCNNPredictor
import torchvision.transforms as T
import cv2
from matplotlib import pyplot as plt
from PIL import Image

import gradio as gr



import sys
sys.path.insert(0, os.path.join(os.path.dirname(__file__), '../', 'src'))
from stacked_hourglass.datasets.imgcropslist import ImgCrops
from combined_model.train_main_image_to_3d_withbreedrel import do_visual_epoch
from combined_model.model_shape_v7 import ModelImageTo3d_withshape_withproj

from configs.barc_cfg_defaults import get_cfg_global_updated



def get_prediction(model, img_path_or_img, confidence=0.5):
    """
    see https://haochen23.github.io/2020/04/object-detection-faster-rcnn.html#.YsMCm4TP3-g
    get_prediction
        parameters:
        - img_path - path of the input image
        - confidence - threshold value for prediction score
        method:
        - Image is obtained from the image path
        - the image is converted to image tensor using PyTorch's Transforms
        - image is passed through the model to get the predictions
        - class, box coordinates are obtained, but only prediction score > threshold
            are chosen.
        
    """
    if isinstance(img_path_or_img, str):
        img = Image.open(img_path_or_img).convert('RGB')
    else:
        img = img_path_or_img
    transform = T.Compose([T.ToTensor()])
    img = transform(img)
    pred = model([img])
    # pred_class = [COCO_INSTANCE_CATEGORY_NAMES[i] for i in list(pred[0]['labels'].numpy())]
    pred_class = list(pred[0]['labels'].numpy())
    pred_boxes = [[(int(i[0]), int(i[1])), (int(i[2]), int(i[3]))] for i in list(pred[0]['boxes'].detach().numpy())]
    pred_score = list(pred[0]['scores'].detach().numpy())
    try:
        pred_t = [pred_score.index(x) for x in pred_score if x>confidence][-1]
        pred_boxes = pred_boxes[:pred_t+1]
        pred_class = pred_class[:pred_t+1]
        return pred_boxes, pred_class, pred_score
    except:
        print('no bounding box with a score that is high enough found! -> work on full image')
        return None, None, None

def detect_object(model, img_path_or_img, confidence=0.5, rect_th=2, text_size=0.5, text_th=1):
    """
    see https://haochen23.github.io/2020/04/object-detection-faster-rcnn.html#.YsMCm4TP3-g
    object_detection_api
        parameters:
        - img_path_or_img - path of the input image
        - confidence - threshold value for prediction score
        - rect_th - thickness of bounding box
        - text_size - size of the class label text
        - text_th - thichness of the text
        method:
        - prediction is obtained from get_prediction method
        - for each prediction, bounding box is drawn and text is written 
            with opencv
        - the final image is displayed
    """
    boxes, pred_cls, pred_scores = get_prediction(model, img_path_or_img, confidence)
    if isinstance(img_path_or_img, str):
        img = cv2.imread(img_path_or_img)
        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    else:
        img = img_path_or_img
    is_first = True
    bbox = None
    if boxes is not None:
        for i in range(len(boxes)):
            cls = pred_cls[i]
            if cls == 18 and bbox is None:
                cv2.rectangle(img, boxes[i][0], boxes[i][1],color=(0, 255, 0), thickness=rect_th)
                # cv2.putText(img, pred_cls[i], boxes[i][0], cv2.FONT_HERSHEY_SIMPLEX, text_size, (0,255,0),thickness=text_th)
                cv2.putText(img, str(pred_scores[i]), boxes[i][0], cv2.FONT_HERSHEY_SIMPLEX, text_size, (0,255,0),thickness=text_th)
                bbox = boxes[i]
    return img, bbox



def run_bbox_inference(input_image):
    model = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True)
    model.eval()
    out_path = os.path.join(cfg.paths.ROOT_OUT_PATH, 'gradio_examples', 'test2.png')
    img, bbox = detect_object(model=model, img_path_or_img=input_image, confidence=0.5)
    fig = plt.figure()   #  plt.figure(figsize=(20,30))
    plt.imsave(out_path, img)
    return img, bbox





def run_barc_inference(input_image, bbox=None):

    # load configs
    cfg = get_cfg_global_updated()

    model_file_complete = os.path.join(cfg.paths.ROOT_CHECKPOINT_PATH, 'barc_complete', 'model_best.pth.tar')  



    # Select the hardware device to use for inference.
    if torch.cuda.is_available() and cfg.device=='cuda':
        device = torch.device('cuda', torch.cuda.current_device())
        # torch.backends.cudnn.benchmark = True
    else:
        device = torch.device('cpu')

    path_model_file_complete = os.path.join(cfg.paths.ROOT_CHECKPOINT_PATH, model_file_complete) 

    # Disable gradient calculations.
    torch.set_grad_enabled(False)

    # prepare complete model
    complete_model = ModelImageTo3d_withshape_withproj(
        num_stage_comb=cfg.params.NUM_STAGE_COMB, num_stage_heads=cfg.params.NUM_STAGE_HEADS, \
        num_stage_heads_pose=cfg.params.NUM_STAGE_HEADS_POSE, trans_sep=cfg.params.TRANS_SEP, \
        arch=cfg.params.ARCH, n_joints=cfg.params.N_JOINTS, n_classes=cfg.params.N_CLASSES, \
        n_keyp=cfg.params.N_KEYP, n_bones=cfg.params.N_BONES, n_betas=cfg.params.N_BETAS, n_betas_limbs=cfg.params.N_BETAS_LIMBS, \
        n_breeds=cfg.params.N_BREEDS, n_z=cfg.params.N_Z, image_size=cfg.params.IMG_SIZE, \
        silh_no_tail=cfg.params.SILH_NO_TAIL, thr_keyp_sc=cfg.params.KP_THRESHOLD, add_z_to_3d_input=cfg.params.ADD_Z_TO_3D_INPUT,
        n_segbps=cfg.params.N_SEGBPS, add_segbps_to_3d_input=cfg.params.ADD_SEGBPS_TO_3D_INPUT, add_partseg=cfg.params.ADD_PARTSEG, n_partseg=cfg.params.N_PARTSEG, \
        fix_flength=cfg.params.FIX_FLENGTH, structure_z_to_betas=cfg.params.STRUCTURE_Z_TO_B, structure_pose_net=cfg.params.STRUCTURE_POSE_NET,
        nf_version=cfg.params.NF_VERSION) 

    # load trained model
    print(path_model_file_complete)
    assert os.path.isfile(path_model_file_complete)
    print('Loading model weights from file: {}'.format(path_model_file_complete))
    checkpoint_complete = torch.load(path_model_file_complete)
    state_dict_complete = checkpoint_complete['state_dict']
    complete_model.load_state_dict(state_dict_complete, strict=False)        
    complete_model = complete_model.to(device)

    save_imgs_path = os.path.join(cfg.paths.ROOT_OUT_PATH, 'gradio_examples')
    if not os.path.exists(save_imgs_path):
        os.makedirs(save_imgs_path)

    input_image_list = [input_image]
    if bbox is not None:
        input_bbox_list = [bbox]
    else:
        input_bbox_list = None
    val_dataset = ImgCrops(image_list=input_image_list, bbox_list=input_bbox_list, dataset_mode='complete')
    test_name_list = val_dataset.test_name_list
    val_loader = DataLoader(val_dataset, batch_size=1, shuffle=False,
                            num_workers=0, pin_memory=True, drop_last=False)   

    # run visual evaluation
    #   remark: take ACC_Joints and DATA_INFO from StanExt as this is the training dataset
    all_results = do_visual_epoch(val_loader, complete_model, device,
                        ImgCrops.DATA_INFO,
                        weight_dict=None,
                        acc_joints=ImgCrops.ACC_JOINTS,
                        save_imgs_path=None, # save_imgs_path,
                        metrics='all', 
                        test_name_list=test_name_list,
                        render_all=cfg.params.RENDER_ALL,
                        pck_thresh=cfg.params.PCK_THRESH, 
                        return_results=True)

    mesh = all_results[0]['mesh_posed']
    result_path = os.path.join(save_imgs_path, test_name_list[0] + '_z')

    mesh.apply_transform([[-1, 0, 0, 0],
                            [0, -1, 0, 0],
                            [0, 0, 1, 1],
                            [0, 0, 0, 1]])
    mesh.export(file_obj=result_path + '.glb')
    result_gltf = result_path + '.glb'
    return [result_gltf, result_gltf]






def run_complete_inference(input_image):

    output_interm_image, output_interm_bbox = run_bbox_inference(input_image.copy())

    print(output_interm_bbox)

    # output_image = run_barc_inference(input_image)
    output_image = run_barc_inference(input_image, output_interm_bbox)

    return output_image




# demo = gr.Interface(run_barc_inference, gr.Image(), "image")
# demo = gr.Interface(run_complete_inference, gr.Image(), "image")



# see: https://huggingface.co/spaces/radames/PIFu-Clothed-Human-Digitization/blob/main/PIFu/spaces.py

description = '''
# BARC

#### Project Page
* https://barc.is.tue.mpg.de/

#### Description
This is a demo for BARC. While BARC is trained on image crops, this demo uses a pretrained Faster-RCNN in order to get bounding boxes for the dogs. 
To see your result you may have to wait a minute or two, please be paitient.

<details>

<summary>More</summary>

#### Citation

```
@inproceedings{BARC:2022,
    title = {BARC}: Learning to Regress {3D} Dog Shape from Images by Exploiting Breed Information,
    author = {Rueegg, Nadine and Zuffi, Silvia and Schindler, Konrad and Black, Michael J.},
    booktitle = {Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR)},
    year = {2022}
}
```

</details>
'''

examples = sorted(glob.glob(os.path.join(os.path.dirname(__file__), '../', 'datasets', 'test_image_crops', '*.jpg')) + glob.glob(os.path.join(os.path.dirname(__file__), '../', 'datasets', 'test_image_crops', '*.png')))  


demo = gr.Interface(
    fn=run_complete_inference,
    description=description,
    # inputs=gr.Image(type="filepath", label="Input Image"),
    inputs=gr.Image(label="Input Image"),
    outputs=[
        gr.Model3D(
            clear_color=[0.0, 0.0, 0.0, 0.0],  label="3D Model"),
        gr.File(label="Download 3D Model")
    ],
    examples=examples,
    thumbnail="barc_thumbnail.png",
    allow_flagging="never",
    cache_examples=True
)



demo.launch(share=True)