app.py

import torch
import torch.nn.functional as F
import logging
import os
import os.path as osp

#os.system('nvidia-smi')

import cupy

import sys
CODE_SPACE=os.path.dirname(os.path.dirname(os.path.abspath(__file__)))

try:
    from mmcv.utils import Config, DictAction
except:
    from mmengine import Config, DictAction
from mono.utils.logger import setup_logger
import glob
from mono.utils.comm import init_env
from mono.model.monodepth_model import get_configured_monodepth_model
from mono.utils.running import load_ckpt
from mono.utils.do_test import transform_test_data_scalecano, get_prediction
from mono.utils.custom_data import load_from_annos, load_data

from mono.utils.avg_meter import MetricAverageMeter
from mono.utils.visualization import save_val_imgs, create_html, save_raw_imgs, save_normal_val_imgs
import cv2
from tqdm import tqdm
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt

from mono.utils.unproj_pcd import reconstruct_pcd, save_point_cloud
from mono.utils.transform import gray_to_colormap
from mono.utils.visualization import vis_surface_normal
import gradio as gr

import fire
from mono.utils.unproj_pcd import reconstruct_pcd, save_point_cloud
from datetime import datetime
import time

#torch.hub.download_url_to_file('https://images.unsplash.com/photo-1437622368342-7a3d73a34c8f', 'turtle.jpg')
#torch.hub.download_url_to_file('https://images.unsplash.com/photo-1519066629447-267fffa62d4b', 'lions.jpg')

cfg_large = Config.fromfile('./mono/configs/HourglassDecoder/vit.raft5.large.py')
model_large = get_configured_monodepth_model(cfg_large, )
model_large, _,  _, _ = load_ckpt('./weight/metric_depth_vit_large_800k.pth', model_large, strict_match=False)
model_large.eval()

cfg_small = Config.fromfile('./mono/configs/HourglassDecoder/vit.raft5.small.py')
model_small = get_configured_monodepth_model(cfg_small, )
model_small, _,  _, _ = load_ckpt('./weight/metric_depth_vit_small_800k.pth', model_small, strict_match=False)
model_small.eval()

device = "cuda"
model_large.to(device)
model_small.to(device)


outputs_dir = "./outs"

def depth_normal(img_path, model_selection="vit-small"):
    if model_selection == "vit-small":
        model = model_small
        cfg = cfg_small
    elif model_selection == "vit-large":
        model = model_large
        cfg = cfg_large

    else:
        raise NotImplementedError
    
    img = Image.open(img_path)

    cv_image = np.array(img) 
    img = cv_image
    img = cv2.cvtColor(cv_image, cv2.COLOR_BGR2RGB)
    intrinsic = [1000.0, 1000.0, img.shape[1]/2, img.shape[0]/2]
    rgb_input, cam_models_stacks, pad, label_scale_factor = transform_test_data_scalecano(img, intrinsic, cfg.data_basic)

    with torch.no_grad():
        pred_depth, pred_depth_scale, scale, output = get_prediction(
                    model = model,
                    input = rgb_input,
                    cam_model = cam_models_stacks,
                    pad_info = pad,
                    scale_info = label_scale_factor,
                    gt_depth = None,
                    normalize_scale = cfg.data_basic.depth_range[1],
                    ori_shape=[img.shape[0], img.shape[1]],
                )

        pred_normal = output['normal_out_list'][0][:, :3, :, :] 
        H, W = pred_normal.shape[2:]
        pred_normal = pred_normal[:, :, pad[0]:H-pad[1], pad[2]:W-pad[3]]

    pred_depth = pred_depth.squeeze().cpu().numpy()
    pred_depth[pred_depth<0] = 0
    pred_color = gray_to_colormap(pred_depth)
    
    ##formatted = (output * 255 / np.max(output)).astype('uint8')
    
    path_output_dir = os.path.splitext(os.path.basename(img_path))[0] + datetime.now().strftime('%Y%m%d-%H%M%S')
    path_output_dir = os.path.join(path_output_dir, outputs_dir)
    os.makedirs(path_output_dir, exist_ok=True)

    name_base = os.path.splitext(os.path.basename(img_path))[0]

    depth_np = pred_depth
    normal_np = torch.nn.functional.interpolate(pred_normal, [img.shape[0], img.shape[1]], mode='bilinear').squeeze().cpu().numpy()
    normal_np = normal_np.transpose(1,2,0)

    pred_normal = pred_normal.squeeze()
    if pred_normal.size(0) == 3:
        pred_normal = pred_normal.permute(1,2,0)
    pred_color_normal = vis_surface_normal(pred_normal)

    depth_path = os.path.join(path_output_dir, f"{name_base}_depth.npy")
    normal_path = os.path.join(path_output_dir, f"{name_base}_normal.npy")

    np.save(normal_path, normal_np)
    np.save(depth_path, depth_np)
    
    ori_w = img.shape[1]
    ori_h = img.shape[0]

    img = Image.fromarray(pred_color)
    #img = img.resize((int(300 * ori_w/ ori_h), 300))

    img_normal = Image.fromarray(pred_color_normal)
    #img_normal = img_normal.resize((int(300 * ori_w/ ori_h), 300))

    return img, img_normal, [depth_path, normal_path]
        
def reconstruction(img_path, files, focal_length, reconstructed_file):
    img = Image.open(img_path)
    cv_image = np.array(img) 
    img = cv_image

    depth_np = np.load(files[0])
    pcd = reconstruct_pcd(depth_np * focal_length / 1000, focal_length, focal_length, img.shape[1]/2, img.shape[0]/2)
    pcd_path = files[0].replace('_depth.npy', '.ply')
    save_point_cloud(pcd.reshape((-1, 3)), img.reshape(-1, 3), pcd_path)
    return [pcd_path]

title = "Metric3D"
description = "Gradio demo for Metric3D which takes in a single image for computing metric depth and surface normal. To use it, simply upload your image, or click one of the examples to load them. Learn more from our paper linked below."
article = "<p style='text-align: center'><a href='https://arxiv.org/pdf/2307.10984.pdf'>Metric3D: Towards Zero-shot Metric 3D Prediction from A Single Image</a> | <a href='https://github.com/YvanYin/Metric3D'>Github Repo</a></p>"

examples = [
    ["files/museum.jpg"],
    ["files/terra.jpg"],
    ["files/underwater.jpg"],
    ["files/venue.jpg"]
]

def run_demo():

    _TITLE = '''Metric3Dv2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation'''
    _DESCRIPTION = description

    with gr.Blocks(title=_TITLE) as demo:
        with gr.Row():
            with gr.Column(scale=1):
                gr.Markdown('# ' + _TITLE)
        gr.Markdown(_DESCRIPTION)
        with gr.Row(variant='panel'):
            with gr.Column(scale=1):
                #input_image = gr.Image(type='pil', label='Original Image')
                input_image = gr.Image(type='filepath', height=300, label='Input image')
                
                example_folder = os.path.join(os.path.dirname(__file__), "./files")
                example_fns = [os.path.join(example_folder, example) for example in os.listdir(example_folder)]
                gr.Examples(
                    examples=example_fns,
                    inputs=[input_image],
                    cache_examples=False,
                    label='Examples (click one of the images below to start)',
                    examples_per_page=30
                )

                model_choice = gr.Dropdown(["vit-small", "vit-large"], label="Model", info="Select a model type",  value="vit-small")
                run_btn = gr.Button('Predict', variant='primary', interactive=True)
        
            with gr.Column(scale=1):
                depth = gr.Image(interactive=False, label="Depth")
                normal = gr.Image(interactive=False, label="Normal")

        with gr.Row():
            files = gr.Files(
                label = "Depth and Normal (numpy)",
                elem_id = "download",
                interactive=False,
            )

        with gr.Row():
            recon_btn = gr.Button('Focal Length Available? If Yes, Enter and Click Here for Metric 3D Reconstruction', variant='primary', interactive=True)
            focal_length = gr.Number(value=1000, label="Focal Length")
        
        with gr.Row():
            reconstructed_file = gr.Files(
                label = "3D pointclouds (plyfile)",
                elem_id = "download",
                interactive=False
            )

        run_btn.click(fn=depth_normal, 
                        inputs=[input_image,
                                model_choice],
                        outputs=[depth, normal, files]
                    )
        recon_btn.click(fn=reconstruction,
                        inputs=[input_image, files, focal_length],
                        outputs=[reconstructed_file]
                        )

        demo.queue().launch(share=True, max_threads=80)


if __name__ == '__main__':
    fire.Fire(run_demo)