app.py

import gradio as gr
import matplotlib.pyplot as plt
import numpy as np
import os
import requests
import spaces
import torch
import torchvision.transforms as T
import types
import albumentations as A
import torch.nn.functional as F

from PIL import Image
from tqdm import tqdm

cmap = plt.get_cmap("tab20")
imagenet_transform = T.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225))

def get_bg_mask(image):
    # detect background based on the four edges
    image = np.array(image)
    if np.all(image[:, 0] == image[0, 0]) and np.all(image[:, -1] == image[0, -1]) \
            and np.all(image[0, :] == image[0, 0]) and np.all(image[-1, :] == image[-1, 0]) \
            and np.all(image[0, 0] == image[0, -1]) and np.all(image[0, 0] == image[-1, 0]) \
            and np.all(image[0, 0] == image[-1, -1]):
        return np.any(image != image[0, 0], -1)
    return np.ones_like(image[:, :, 0], dtype=bool)

    
def download_image(url, save_path):
    response = requests.get(url)
    with open(save_path, 'wb') as file:
        file.write(response.content)


def process_image(image, res, patch_size, decimation=4):
    image = torch.from_numpy(np.array(image) / 255.).float().permute(2, 0, 1).to(device)

    tgt_size = (int(image.shape[-2] * res / image.shape[-1]), res)
    if image.shape[-2] > image.shape[-1]:
        tgt_size = (res, int(image.shape[-1] * res / image.shape[-2]))
    
    patch_h, patch_w = tgt_size[0] // decimation, tgt_size[1] // decimation
    image_resized = T.functional.resize(image, (patch_h * patch_size, patch_w * patch_size))
    
    image_resized = imagenet_transform(image_resized)

    return image_resized

def generate_grid(x, y, stride):
    x_coords = np.arange(0, x, grid_stride)
    y_coords = np.arange(0, y, grid_stride)

    x_mesh, y_mesh = np.meshgrid(x_coords, y_coords)
    kp = np.column_stack((x_mesh.ravel(), y_mesh.ravel())).astype(float)
    return kp

def pca(feat, pca_dim=3):
    feat_flattened = feat
    mean = torch.mean(feat_flattened, dim=0)
    centered_features = feat_flattened - mean
    U, S, V = torch.pca_lowrank(centered_features, q=pca_dim)
    reduced_features = torch.matmul(centered_features, V[:, :pca_dim])
    return reduced_features
    
def co_pca(feat1, feat2, pca_dim=3):
    co_feats = torch.cat((feat1.reshape(-1, feat1.shape[-1]), feat2.reshape(-1, feat2.shape[-1])), dim=0)
    feats = pca(co_feats)
    feat1_pca = feats[:feat1.shape[0]*feat1.shape[1]].reshape(feat1.shape[0], feat1.shape[1], -1)
    feat2_pca = feats[feat1.shape[0]*feat1.shape[1]:].reshape(feat2.shape[0], feat2.shape[1], -1)
    return feat1_pca, feat2_pca
    
    
def draw_correspondence(feat1, feat2, color1, mask1, mask2):
    original_mask2_shape = mask2.shape
    mask1, mask2 = mask1.reshape(-1), mask2.reshape(-1)
    distances = torch.cdist(feat1.reshape(-1, feat1.shape[-1])[mask1], feat2.reshape(-1, feat2.shape[-1])[mask2])
    nearest = torch.argmin(distances, dim=0)
    color2 = torch.zeros((mask2.shape[0], 3,)).to(device)
    color2[mask2] = color1.reshape(-1, 3)[mask1][nearest]
    color2 = color2.reshape(*original_mask2_shape, 3)
    return color2
    
def load_model(options):
    original_models = {}
    fine_models = {}
    for option in tqdm(options):
        print('Please wait ...')
        print('loading weights of ', option)
        fine_models[option] = torch.hub.load(".", model_card[option], source='local').to(device)
        original_models[option] = torch.hub.load(repo_or_dir="facebookresearch/dinov2", model=fine_models[option].backbone_name).eval().to(device)
    print('Done! Now play the demo :)')
    return original_models, fine_models
        
if __name__ == "__main__":

    if torch.cuda.is_available():
        device = torch.device('cuda')
    else:
        device = torch.device('cpu')

    print("device: ")
    print(device)

    example_dir = "examples"

    os.makedirs(example_dir, exist_ok=True)
        
    image_input1 = gr.Image(label="Choose an image:",
                           height=500,
                           type="pil",
                           image_mode='RGB',
                           sources=['upload', 'webcam', 'clipboard']
                           )
    image_input2 = gr.Image(label="Choose another image:",
                            height=500,
                            type="pil",
                            image_mode='RGB',
                            sources=['upload', 'webcam', 'clipboard']
                            )

    options = ['DINOv2-Base']
    model_option = gr.Radio(options, value="DINOv2-Base", label='Choose a 2D foundation model')
    
    model_card = {
        "DINOv2-Base": "dinov2_base",
    }


    os.environ['TORCH_HOME'] = '/tmp/.cache'
    # os.environ['GRADIO_EXAMPLES_CACHE'] = '/tmp/gradio_cache'

    # Pre-load all models
    original_models, fine_models = load_model(options)

    @spaces.GPU
    def main(image1, image2, model_option, kmeans_num):
        if image1 is None or image2 is None:
            return None
        # Select model
        original_model = original_models[model_option]
        fine_model = fine_models[model_option]
        
        images_resized = [process_image(image, 640, 14, decimation=8) for image in [image1, image2]]
        masks = [torch.from_numpy(get_bg_mask(image)).to(device) for image in [image1, image2]]
        feat_shapes = [(images_resized[0].shape[-2] // 14, images_resized[0].shape[-1] // 14),
                       (images_resized[1].shape[-2] // 14, images_resized[1].shape[-1] // 14)]

        masks_resized = [T.functional.resize(mask.float()[None], feat_shape,
                                            interpolation=T.functional.InterpolationMode.NEAREST_EXACT)[0]
                         for mask, feat_shape in zip(masks, feat_shapes)]
        
        with torch.no_grad():
            original_feats = [original_model.forward_features(image[None])['x_norm_patchtokens'].reshape(*feat_shape, -1)
                              for image, feat_shape in zip(images_resized, feat_shapes)]
            original_feats = [F.normalize(feat, p=2, dim=-1) for feat in original_feats]
            
            original_color1 = torch.zeros((original_feats[0].shape[0] * original_feats[0].shape[1], 3,)).to(device)
            color = pca((original_feats[0][masks_resized[0] > 0]), 3)
            color = (color - color.min()) / (color.max() - color.min())
            original_color1[masks_resized[0].reshape(-1) > 0] = color
            original_color1 = original_color1.reshape(*original_feats[0].shape[:2], 3)
            
            original_color2 = draw_correspondence(original_feats[0], original_feats[1], original_color1, 
                                                  masks_resized[0] > 0, masks_resized[1] > 0)
            
            fine_feats = [fine_model.dinov2.forward_features(image[None])['x_norm_patchtokens'].reshape(*feat_shape, -1)
                              for image, feat_shape in zip(images_resized, feat_shapes)]
            fine_feats = [fine_model.refine_conv(feat[None].permute(0, 3, 1, 2)).permute(0, 2, 3, 1)[0] for feat in fine_feats]
            fine_feats = [F.normalize(feat, p=2, dim=-1) for feat in fine_feats]
            fine_color2 = draw_correspondence(fine_feats[0], fine_feats[1], original_color1, 
                                              masks_resized[0] > 0, masks_resized[1] > 0)

        fig, ax = plt.subplots(2, 2, squeeze=False)
        ax[0][0].imshow(original_color1.cpu().numpy())
        ax[0][1].text(-0.1, 0.5, "Original " + model_option, fontsize=7, rotation=90, va='center', transform=ax[0][1].transAxes)
        ax[0][1].imshow(original_color2.cpu().numpy())
        
        # ax[1][0].imshow(fine_color1.cpu().numpy())
        ax[1][1].text(-0.1, 0.5, "Finetuned " + model_option, fontsize=7, rotation=90, va='center', transform=ax[1][1].transAxes)
        ax[1][1].imshow(fine_color2.cpu().numpy())
        for xx in ax:
            for x in xx:
                x.xaxis.set_major_formatter(plt.NullFormatter())
                x.yaxis.set_major_formatter(plt.NullFormatter())
                x.set_xticks([])
                x.set_yticks([])
                x.axis('off')

        plt.tight_layout()
        plt.close(fig)
        return fig

    
    demo = gr.Interface(
        title="<center>  \
        <h1>3DCorrEnhance</h1> \
        <h2>Multiview Equivariance Improves 3D Correspondence Understanding with Minimal Feature Finetuning</h2> \
        <h2>ICLR 2025</h2> \
        </center>",
        description="",
        fn=main, 
        inputs=[image_input1, image_input2, model_option],
        outputs="plot",
        examples=[
            ["examples/objs/1-1.png", "examples/objs/1-2.png", "DINOv2-Base"],
            ["examples/scenes/1-1.jpg", "examples/scenes/1-2.jpg", "DINOv2-Base"],
            ["examples/scenes/2-1.jpg", "examples/scenes/2-2.jpg", "DINOv2-Base"],
            ],
        cache_examples=True)
    demo.launch()