|
| 1 | +import os |
| 2 | +import torch |
| 3 | +import torchvision.transforms as T |
| 4 | +import matplotlib.pyplot as plt |
| 5 | +import numpy as np |
| 6 | +from PIL import Image |
| 7 | +from sklearn.decomposition import PCA |
| 8 | +import argparse |
| 9 | +from romatch.models.transformer import vit_base |
| 10 | + |
| 11 | +device = "cuda" if torch.cuda.is_available() else "cpu" |
| 12 | + |
| 13 | + |
| 14 | +# -------------------- 可视化函数 -------------------- |
| 15 | +def vis_feat_map(features, patch_h, patch_w, resize_hw=(560, 560)): |
| 16 | + features = features.reshape(patch_h * patch_w, -1) |
| 17 | + pca = PCA(n_components=3) |
| 18 | + pca_feats = pca.fit_transform(features) |
| 19 | + pca_feats = (pca_feats - pca_feats.mean(0)) / (pca_feats.std(0) + 1e-5) |
| 20 | + pca_feats = np.clip(pca_feats * 0.5 + 0.5, 0, 1) |
| 21 | + img = pca_feats.reshape(patch_h, patch_w, 3) |
| 22 | + img = (img * 255).astype(np.uint8) |
| 23 | + img = Image.fromarray(img) |
| 24 | + return img.resize(resize_hw, Image.BICUBIC) |
| 25 | + |
| 26 | + |
| 27 | + |
| 28 | +def save_all_visualizations( |
| 29 | + feat_dino, feat_fit3d, feat_L2M, |
| 30 | + patch_h, patch_w, base_name, save_dir, original_image=None |
| 31 | +): |
| 32 | + os.makedirs(save_dir, exist_ok=True) |
| 33 | + |
| 34 | + # 单图保存 |
| 35 | + img_dino = vis_feat_map(feat_dino, patch_h, patch_w) |
| 36 | + img_fit3d = vis_feat_map(feat_fit3d, patch_h, patch_w) |
| 37 | + img_L2M = vis_feat_map(feat_L2M, patch_h, patch_w) |
| 38 | + |
| 39 | + img_dino.save(os.path.join(save_dir, f"{base_name}_dino.png")) |
| 40 | + img_fit3d.save(os.path.join(save_dir, f"{base_name}_fit3d_fit3d.png")) |
| 41 | + img_L2M.save(os.path.join(save_dir, f"{base_name}_L2M.png")) |
| 42 | + |
| 43 | + # 拼图（含原图） |
| 44 | + fig, ax = plt.subplots(1, 4, figsize=(16, 4)) |
| 45 | + for a, im, title in zip( |
| 46 | + ax, |
| 47 | + [original_image, img_dino, img_fit3d, img_L2M], |
| 48 | + ["Original", "DINOv2", "Fit3D", "L2M (Ours)"] |
| 49 | + ): |
| 50 | + a.imshow(im) |
| 51 | + a.set_title(title, fontsize=12) |
| 52 | + a.axis("off") |
| 53 | + plt.tight_layout() |
| 54 | + plt.savefig(os.path.join(save_dir, f"{base_name}_compare.png")) |
| 55 | + plt.close() |
| 56 | + |
| 57 | + |
| 58 | +# -------------------- 特征提取函数 -------------------- |
| 59 | +def extract_features(model, image_tensor): |
| 60 | + with torch.no_grad(): |
| 61 | + return model.forward_features(image_tensor)["x_norm_patchtokens"].squeeze(0).cpu().numpy() |
| 62 | + |
| 63 | + |
| 64 | +# -------------------- 主脚本 -------------------- |
| 65 | +def main(args): |
| 66 | + os.makedirs(args.save_dir, exist_ok=True) |
| 67 | + |
| 68 | + patch_h, patch_w = 37, 37 |
| 69 | + img_size = patch_h * 14 # = 560 |
| 70 | + feat_dim = 768 |
| 71 | + |
| 72 | + transform = T.Compose([ |
| 73 | + T.GaussianBlur(9, sigma=(0.1, 2.0)), |
| 74 | + T.Resize((img_size, img_size)), |
| 75 | + T.CenterCrop((img_size, img_size)), |
| 76 | + T.ToTensor(), |
| 77 | + T.Normalize(mean=(0.485, 0.456, 0.406), |
| 78 | + std=(0.229, 0.224, 0.225)), |
| 79 | + ]) |
| 80 | + |
| 81 | + # 初始化模型 |
| 82 | + vit_kwargs = dict( |
| 83 | + img_size=img_size, |
| 84 | + patch_size=14, |
| 85 | + init_values=1.0, |
| 86 | + ffn_layer="mlp", |
| 87 | + block_chunks=0 |
| 88 | + ) |
| 89 | + |
| 90 | + # DINOv2 |
| 91 | + dino = vit_base(**vit_kwargs).eval().to(device) |
| 92 | + dino_ckpt_raw = torch.load(args.ckpt_dino, map_location="cpu") |
| 93 | + dino_ckpt = {k.replace("model.", ""): v for k, v in dino_ckpt_raw.items()} |
| 94 | + dino.load_state_dict(dino_ckpt, strict=False) |
| 95 | + |
| 96 | + # Fit3D |
| 97 | + fit3d = vit_base(**vit_kwargs).eval().to(device) |
| 98 | + fit3d_ckpt_raw = torch.load(args.ckpt_fit3d, map_location="cpu")["model"] |
| 99 | + fit3d_ckpt = {k.replace("model.", ""): v for k, v in fit3d_ckpt_raw.items()} |
| 100 | + fit3d.load_state_dict(fit3d_ckpt, strict=False) |
| 101 | + |
| 102 | + # L2M (Ours) |
| 103 | + L2M = vit_base(**vit_kwargs).eval().to(device) |
| 104 | + L2M_ckpt = torch.load(args.ckpt_L2M, map_location="cpu") |
| 105 | + L2M.load_state_dict(L2M_ckpt, strict=False) |
| 106 | + |
| 107 | + for i, img_path in enumerate(args.img_paths): |
| 108 | + img = Image.open(img_path).convert("RGB") |
| 109 | + x = transform(img).unsqueeze(0).to(device) |
| 110 | + |
| 111 | + # 提取特征 |
| 112 | + feat_dino = extract_features(dino, x) |
| 113 | + feat_fit3d = extract_features(fit3d, x) |
| 114 | + feat_L2M = extract_features(L2M, x) |
| 115 | + |
| 116 | + # 保存图 |
| 117 | + base_name = os.path.splitext(os.path.basename(img_path))[0] |
| 118 | + save_all_visualizations( |
| 119 | + feat_dino, feat_fit3d, feat_L2M, |
| 120 | + patch_h, patch_w, base_name, args.save_dir, |
| 121 | + original_image=img |
| 122 | + ) |
| 123 | + |
| 124 | + |
| 125 | + print(f"[{i+1}/{len(args.img_paths)}] Saved visualizations for {img_path}") |
| 126 | + |
| 127 | + |
| 128 | +if __name__ == "__main__": |
| 129 | + parser = argparse.ArgumentParser() |
| 130 | + parser.add_argument( |
| 131 | + "--img_paths", |
| 132 | + nargs="+", |
| 133 | + default=[ |
| 134 | + "assets/sacre_coeur_A.jpg", |
| 135 | + "assets/sacre_coeur_B.jpg" |
| 136 | + ], |
| 137 | + help="List of image paths" |
| 138 | + ) |
| 139 | + parser.add_argument( |
| 140 | + "--ckpt_fit3d", |
| 141 | + default="ckpts/fit3d.pth", |
| 142 | + help="Original Fit3D checkpoint" |
| 143 | + ) |
| 144 | + parser.add_argument( |
| 145 | + "--ckpt_L2M", |
| 146 | + default="ckpts/output_20250629/l2m_vit_base.pth", |
| 147 | + help="L2M Fit3D checkpoint" |
| 148 | + ) |
| 149 | + parser.add_argument( |
| 150 | + "--ckpt_dino", |
| 151 | + default="ckpts/dinov2.pth", |
| 152 | + help="dino checkpoint" |
| 153 | + ) |
| 154 | + parser.add_argument( |
| 155 | + "--save_dir", |
| 156 | + default="outputs_vis_feat", |
| 157 | + help="Directory to save visualizations" |
| 158 | + ) |
| 159 | + args = parser.parse_args() |
| 160 | + main(args) |
0 commit comments