import os import torch import os import pandas as pd import numpy as np from PIL import Image from tqdm import tqdm import torch.nn.functional as F from extractor_sd import load_model, load_sd_backbone, process_features_and_mask, get_mask from open_clip.transform import ResizeMaxSize,_convert_to_rgb from torchvision.transforms import ToTensor from third_party.utils.utils_correspondence import co_pca, resize, find_nearest_patchs, find_nearest_patchs_replace import matplotlib.pyplot as plt import sys from detectron2.data import transforms as T from extractor_dino import ViTExtractor from sklearn.decomposition import PCA as sklearnPCA import math from sklearn.cluster import KMeans from scipy.optimize import linear_sum_assignment from torchvision import transforms def preprocess_pil(pil_image): prep = transforms.Compose([ transforms.ToTensor(), transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)) ]) prep_img = prep(pil_image)[None, ...] return prep_img MASK = True VER = "v1-5" PCA = False CO_PCA = True PCA_DIMS = [256, 256, 256] SIZE =960 EDGE_PAD = False FUSE_DINO = 1 ONLY_DINO = 0 MODEL_SIZE = 'base' DRAW=1 TEXT_INPUT = False SEED = 42 TIMESTEP = 100 DIST = 'l2' if FUSE_DINO and not ONLY_DINO else 'cos' if ONLY_DINO: FUSE_DINO = True np.random.seed(SEED) torch.manual_seed(SEED) torch.cuda.manual_seed(SEED) torch.backends.cudnn.benchmark = True sd_transform=transforms.Compose([ ResizeMaxSize(960, fill=0), _convert_to_rgb, ToTensor(), ]) model = load_sd_backbone(diffusion_ver=VER, image_size=SIZE, num_timesteps=TIMESTEP, decoder_only=False) odise, aug = load_model(diffusion_ver=VER, image_size=SIZE, num_timesteps=TIMESTEP,decoder_only=False) def compute_pair_feature(model, aug, save_path, files, category, mask=False, dist='cos', real_size=960): img_size = 840 stride = 14 device = 'cuda' if torch.cuda.is_available() else 'cpu' # extractor=torch.hub.load('/mnt/SSD8T/home/wjj/.cache/torch/hub/facebookresearch_dinov2_main', 'dinov2_vitb14', source='local').to(device).eval() extractor = ViTExtractor('dinov2_vitb14', 14, device=device) patch_size = 14 num_patches = int(patch_size / stride * (img_size // patch_size - 1) + 1) # Load image 1 # img1_input = resize(img1, real_size, resize=True, to_pil=True, edge=EDGE_PAD) img1 = Image.open(files[0]) sd_input1=sd_transform(img1).to(device).unsqueeze(0) img1 = resize(img1, img_size, resize=True, to_pil=True, edge=EDGE_PAD) # Load image 2 # img2_input = resize(img2, real_size, resize=True, to_pil=True, edge=EDGE_PAD) img2 = Image.open(files[1]) sd_input2 = sd_transform(img2).to(device).unsqueeze(0) img2 = resize(img2, img_size, resize=True, to_pil=True, edge=EDGE_PAD) result = [] with torch.no_grad(): # features1 = process_features_and_mask(model, aug, img1_input, input_text=input_text, mask=False, raw=True) # features2 = process_features_and_mask(model, aug, img2_input, input_text=input_text, mask=False, raw=True) features1=model(sd_input1,raw=True) features2=model(sd_input2,raw=True) processed_features1, processed_features2 = co_pca(features1, features2, PCA_DIMS) img1_desc = processed_features1.reshape(1, 1, -1, num_patches**2).permute(0,1,3,2) img2_desc = processed_features2.reshape(1, 1, -1, num_patches**2).permute(0,1,3,2) img1_batch = preprocess_pil(img1).to(device) # img1_desc_dino = extractor.get_intermediate_layers(img1_batch)[0].unsqueeze(1) img1_desc_dino = extractor.extract_descriptors(img1_batch, 11, 'token') img2_batch = preprocess_pil(img2).to(device) # img2_desc_dino = extractor.get_intermediate_layers(img2_batch)[0].unsqueeze(1) img2_desc_dino = extractor.extract_descriptors(img2_batch, 11, 'token') img1_desc = img1_desc / img1_desc.norm(dim=-1, keepdim=True) img2_desc = img2_desc / img2_desc.norm(dim=-1, keepdim=True) img1_desc_dino = img1_desc_dino / img1_desc_dino.norm(dim=-1, keepdim=True) img2_desc_dino = img2_desc_dino / img2_desc_dino.norm(dim=-1, keepdim=True) img1_desc = torch.cat((img1_desc, img1_desc_dino), dim=-1) img2_desc = torch.cat((img2_desc, img2_desc_dino), dim=-1) if DRAW: mask1 = get_mask(odise, aug, img1, category[0]) mask2 = get_mask(odise, aug, img2, category[-1]) img1_desc_reshaped = img1_desc.permute(0,1,3,2).reshape(-1, img1_desc.shape[-1], num_patches, num_patches) img2_desc_reshaped = img2_desc.permute(0,1,3,2).reshape(-1, img2_desc.shape[-1], num_patches, num_patches) trg_dense_output, src_color_map = find_nearest_patchs(mask2, mask1, img2, img1, img2_desc_reshaped, img1_desc_reshaped, mask=mask) if not os.path.exists(f'{save_path}/{category[0]}'): os.makedirs(f'{save_path}/{category[0]}') fig_colormap, (ax1, ax2) = plt.subplots(1, 2, figsize=(16, 8)) ax1.axis('off') ax2.axis('off') ax1.imshow(src_color_map) ax2.imshow(trg_dense_output) fig_colormap.savefig(f'{save_path}/{category[0]}/_colormap.png') plt.close(fig_colormap) img1_desc_reshaped = img1_desc.permute(0,1,3,2).reshape(-1, img1_desc.shape[-1], num_patches, num_patches) img2_desc_reshaped = img2_desc.permute(0,1,3,2).reshape(-1, img2_desc.shape[-1], num_patches, num_patches) trg_dense_output, src_color_map = find_nearest_patchs_replace(mask2, mask1, img2, img1, img2_desc_reshaped, img1_desc_reshaped, mask=mask, resolution=156) if not os.path.exists(f'{save_path}/{category[0]}'): os.makedirs(f'{save_path}/{category[0]}') fig_colormap, (ax1, ax2) = plt.subplots(1, 2, figsize=(16, 8)) ax1.axis('off') ax2.axis('off') ax1.imshow(src_color_map) ax2.imshow(trg_dense_output) fig_colormap.savefig(f'{save_path}/{category[0]}/_swap.png') plt.close(fig_colormap) result.append([img1_desc.cpu(), img2_desc.cpu(), mask1.cpu(), mask2.cpu()]) return result def vis_pca_mask(result,save_path): # PCA visualization mask version for (feature1,feature2,mask1,mask2) in result: # feature1 shape (1,1,3600,768*2) # feature2 shape (1,1,3600,768*2) num_patches = int(math.sqrt(feature1.shape[-2])) # pca the concatenated feature to 3 dimensions feature1 = feature1.squeeze() # shape (3600,768*2) feature2 = feature2.squeeze() # shape (3600,768*2) chennel_dim = feature1.shape[-1] # resize back src_feature_reshaped = feature1.squeeze().permute(1,0).reshape(-1,num_patches,num_patches).cuda() tgt_feature_reshaped = feature2.squeeze().permute(1,0).reshape(-1,num_patches,num_patches).cuda() resized_src_mask = F.interpolate(mask1.unsqueeze(0).unsqueeze(0), size=(num_patches, num_patches), mode='nearest').squeeze().cuda() resized_tgt_mask = F.interpolate(mask2.unsqueeze(0).unsqueeze(0), size=(num_patches, num_patches), mode='nearest').squeeze().cuda() src_feature_upsampled = src_feature_reshaped * resized_src_mask.repeat(src_feature_reshaped.shape[0],1,1) tgt_feature_upsampled = tgt_feature_reshaped * resized_tgt_mask.repeat(src_feature_reshaped.shape[0],1,1) feature1=src_feature_upsampled.reshape(chennel_dim,-1).permute(1,0) feature2=tgt_feature_upsampled.reshape(chennel_dim,-1).permute(1,0) n_components=4 # the first component is to seperate the object from the background pca = sklearnPCA(n_components=n_components) feature1_n_feature2 = torch.cat((feature1,feature2),dim=0) # shape (7200,768*2) feature1_n_feature2 = pca.fit_transform(feature1_n_feature2.cpu().numpy()) # shape (7200,3) feature1 = feature1_n_feature2[:feature1.shape[0],:] # shape (3600,3) feature2 = feature1_n_feature2[feature1.shape[0]:,:] # shape (3600,3) fig, axes = plt.subplots(4, 2, figsize=(10, 14)) for show_channel in range(n_components): if show_channel==0: continue # min max normalize the feature map feature1[:, show_channel] = (feature1[:, show_channel] - feature1[:, show_channel].min()) / (feature1[:, show_channel].max() - feature1[:, show_channel].min()) feature2[:, show_channel] = (feature2[:, show_channel] - feature2[:, show_channel].min()) / (feature2[:, show_channel].max() - feature2[:, show_channel].min()) feature1_first_channel = feature1[:, show_channel].reshape(num_patches,num_patches) feature2_first_channel = feature2[:, show_channel].reshape(num_patches,num_patches) axes[show_channel-1, 0].imshow(feature1_first_channel) axes[show_channel-1, 0].axis('off') axes[show_channel-1, 1].imshow(feature2_first_channel) axes[show_channel-1, 1].axis('off') axes[show_channel-1, 0].set_title('Feature 1 - Channel {}'.format(show_channel ), fontsize=14) axes[show_channel-1, 1].set_title('Feature 2 - Channel {}'.format(show_channel ), fontsize=14) feature1_resized = feature1[:, 1:4].reshape(num_patches,num_patches, 3) feature2_resized = feature2[:, 1:4].reshape(num_patches,num_patches, 3) axes[3, 0].imshow(feature1_resized) axes[3, 0].axis('off') axes[3, 1].imshow(feature2_resized) axes[3, 1].axis('off') axes[3, 0].set_title('Feature 1 - All Channels', fontsize=14) axes[3, 1].set_title('Feature 2 - All Channels', fontsize=14) plt.tight_layout() plt.show() fig.savefig(save_path+'/masked_pca.png', dpi=300) def vis_pca(result,save_path,src_img_path,trg_img_path): # PCA visualization for (feature1,feature2,mask1,mask2) in result: # feature1 shape (1,1,3600,768*2) # feature2 shape (1,1,3600,768*2) num_patches=int(math.sqrt(feature1.shape[2])) # pca the concatenated feature to 3 dimensions feature1 = feature1.squeeze() # shape (3600,768*2) feature2 = feature2.squeeze() # shape (3600,768*2) chennel_dim = feature1.shape[-1] # resize back h1, w1 = Image.open(src_img_path).size scale_h1 = h1/num_patches scale_w1 = w1/num_patches if scale_h1 > scale_w1: scale = scale_h1 scaled_w = int(w1/scale) feature1 = feature1.reshape(num_patches,num_patches,chennel_dim) feature1_uncropped=feature1[(num_patches-scaled_w)//2:num_patches-(num_patches-scaled_w)//2,:,:] else: scale = scale_w1 scaled_h = int(h1/scale) feature1 = feature1.reshape(num_patches,num_patches,chennel_dim) feature1_uncropped=feature1[:,(num_patches-scaled_h)//2:num_patches-(num_patches-scaled_h)//2,:] h2, w2 = Image.open(trg_img_path).size scale_h2 = h2/num_patches scale_w2 = w2/num_patches if scale_h2 > scale_w2: scale = scale_h2 scaled_w = int(w2/scale) feature2 = feature2.reshape(num_patches,num_patches,chennel_dim) feature2_uncropped=feature2[(num_patches-scaled_w)//2:num_patches-(num_patches-scaled_w)//2,:,:] else: scale = scale_w2 scaled_h = int(h2/scale) feature2 = feature2.reshape(num_patches,num_patches,chennel_dim) feature2_uncropped=feature2[:,(num_patches-scaled_h)//2:num_patches-(num_patches-scaled_h)//2,:] f1_shape=feature1_uncropped.shape[:2] f2_shape=feature2_uncropped.shape[:2] feature1 = feature1_uncropped.reshape(f1_shape[0]*f1_shape[1],chennel_dim) feature2 = feature2_uncropped.reshape(f2_shape[0]*f2_shape[1],chennel_dim) n_components=3 pca = sklearnPCA(n_components=n_components) feature1_n_feature2 = torch.cat((feature1,feature2),dim=0) # shape (7200,768*2) feature1_n_feature2 = pca.fit_transform(feature1_n_feature2.cpu().numpy()) # shape (7200,3) feature1 = feature1_n_feature2[:feature1.shape[0],:] # shape (3600,3) feature2 = feature1_n_feature2[feature1.shape[0]:,:] # shape (3600,3) fig, axes = plt.subplots(4, 2, figsize=(10, 14)) for show_channel in range(n_components): # min max normalize the feature map feature1[:, show_channel] = (feature1[:, show_channel] - feature1[:, show_channel].min()) / (feature1[:, show_channel].max() - feature1[:, show_channel].min()) feature2[:, show_channel] = (feature2[:, show_channel] - feature2[:, show_channel].min()) / (feature2[:, show_channel].max() - feature2[:, show_channel].min()) feature1_first_channel = feature1[:, show_channel].reshape(f1_shape[0], f1_shape[1]) feature2_first_channel = feature2[:, show_channel].reshape(f2_shape[0], f2_shape[1]) axes[show_channel, 0].imshow(feature1_first_channel) axes[show_channel, 0].axis('off') axes[show_channel, 1].imshow(feature2_first_channel) axes[show_channel, 1].axis('off') axes[show_channel, 0].set_title('Feature 1 - Channel {}'.format(show_channel + 1), fontsize=14) axes[show_channel, 1].set_title('Feature 2 - Channel {}'.format(show_channel + 1), fontsize=14) feature1_resized = feature1[:, :3].reshape(f1_shape[0], f1_shape[1], 3) feature2_resized = feature2[:, :3].reshape(f2_shape[0], f2_shape[1], 3) axes[3, 0].imshow(feature1_resized) axes[3, 0].axis('off') axes[3, 1].imshow(feature2_resized) axes[3, 1].axis('off') axes[3, 0].set_title('Feature 1 - All Channels', fontsize=14) axes[3, 1].set_title('Feature 2 - All Channels', fontsize=14) plt.tight_layout() plt.show() fig.savefig(save_path+'/pca.png', dpi=300) def perform_clustering(features, n_clusters=10): # Normalize features features = F.normalize(features, p=2, dim=1) # Convert the features to float32 features = features.cpu().detach().numpy().astype('float32') # Initialize a k-means clustering index with the desired number of clusters kmeans = KMeans(n_clusters=n_clusters, random_state=0) # Train the k-means index with the features kmeans.fit(features) # Assign the features to their nearest cluster labels = kmeans.predict(features) return labels def cluster_and_match(result, save_path, n_clusters=6): for (feature1,feature2,mask1,mask2) in result: # feature1 shape (1,1,3600,768*2) num_patches = int(math.sqrt(feature1.shape[-2])) # pca the concatenated feature to 3 dimensions feature1 = feature1.squeeze() # shape (3600,768*2) feature2 = feature2.squeeze() # shape (3600,768*2) chennel_dim = feature1.shape[-1] # resize back src_feature_reshaped = feature1.squeeze().permute(1,0).reshape(-1,num_patches,num_patches).cuda() tgt_feature_reshaped = feature2.squeeze().permute(1,0).reshape(-1,num_patches,num_patches).cuda() resized_src_mask = F.interpolate(mask1.unsqueeze(0).unsqueeze(0), size=(num_patches, num_patches), mode='nearest').squeeze().cuda() resized_tgt_mask = F.interpolate(mask2.unsqueeze(0).unsqueeze(0), size=(num_patches, num_patches), mode='nearest').squeeze().cuda() src_feature_upsampled = src_feature_reshaped * resized_src_mask.repeat(src_feature_reshaped.shape[0],1,1) tgt_feature_upsampled = tgt_feature_reshaped * resized_tgt_mask.repeat(src_feature_reshaped.shape[0],1,1) feature1=src_feature_upsampled.unsqueeze(0) feature2=tgt_feature_upsampled.unsqueeze(0) w1, h1 = feature1.shape[2], feature1.shape[3] w2, h2 = feature2.shape[2], feature2.shape[3] features1_2d = feature1.reshape(feature1.shape[1], -1).permute(1, 0) features2_2d = feature2.reshape(feature2.shape[1], -1).permute(1, 0) labels_img1 = perform_clustering(features1_2d, n_clusters) labels_img2 = perform_clustering(features2_2d, n_clusters) cluster_means_img1 = [features1_2d.cpu().detach().numpy()[labels_img1 == i].mean(axis=0) for i in range(n_clusters)] cluster_means_img2 = [features2_2d.cpu().detach().numpy()[labels_img2 == i].mean(axis=0) for i in range(n_clusters)] distances = np.linalg.norm(np.expand_dims(cluster_means_img1, axis=1) - np.expand_dims(cluster_means_img2, axis=0), axis=-1) # Use Hungarian algorithm to find the optimal bijective mapping row_ind, col_ind = linear_sum_assignment(distances) relabeled_img2 = np.zeros_like(labels_img2) for i, match in zip(row_ind, col_ind): relabeled_img2[labels_img2 == match] = i labels_img1 = labels_img1.reshape(w1, h1) relabeled_img2 = relabeled_img2.reshape(w2, h2) fig, axs = plt.subplots(1, 2, figsize=(10, 5)) # Plot the results ax_img1 = axs[0] axs[0].axis('off') ax_img1.imshow(labels_img1, cmap='tab20') ax_img2 = axs[1] axs[1].axis('off') ax_img2.imshow(relabeled_img2, cmap='tab20') plt.tight_layout() plt.show() fig.savefig(save_path+'/clustering.png', dpi=300) def process_images(src_img_path,trg_img_path): categories = [['dog'], ['dog']] files = [src_img_path, trg_img_path] save_path = './results_vis' + f'/{trg_img_path.split("/")[-1].split(".")[0]}_{src_img_path.split("/")[-1].split(".")[0]}' result = compute_pair_feature(model, aug, save_path, files, mask=MASK, category=categories, dist=DIST) if MASK: vis_pca_mask(result, save_path) cluster_and_match(result, save_path) if 'Anno' not in src_img_path: vis_pca(result, save_path,src_img_path,trg_img_path) return result src_img_path = "/mnt/SSD8T/home/wjj/code/sd-dino/data/images/dog_00.jpg" trg_img_path = "/mnt/SSD8T/home/wjj/code/sd-dino/data/images/dog_59.jpg" result = process_images(src_img_path, trg_img_path)