Upload biplet-asmk-mast3r-ps2-gs-kg-32.ipynb

biplet-asmk-mast3r-ps2-gs-kg-32.ipynb

@@ -0,0 +1 @@
+{"metadata":{"kernelspec":{"name":"python3","display_name":"Python 3","language":"python"},"language_info":{"name":"python","version":"3.12.12","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"colab":{"provenance":[],"gpuType":"T4"},"accelerator":"GPU","kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceId":14571475,"sourceType":"datasetVersion","datasetId":1429416}],"dockerImageVersionId":31260,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# **biplet-asmk-mast3r-ps2-gs-kg** \n\n","metadata":{"id":"qDQLX3PArmh8"}},{"cell_type":"markdown","source":"https://www.kaggle.com/code/stpeteishii/dino-mast3r-gs-kg-34","metadata":{}},{"cell_type":"markdown","source":"v.32 全面見直し","metadata":{}},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =====================================================================\n# CELL 1: Install Dependencies\n# =====================================================================\n!pip install roma einops timm huggingface_hub\n!pip install opencv-python pillow tqdm pyaml cython plyfile\n!pip install pycolmap trimesh\n!pip uninstall -y numpy scipy\n!pip install numpy==1.26.4 scipy==1.11.4\nbreak","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =====================================================================\n# CELL 2: Restart Kernel (Run this after Cell 1)\n# =====================================================================\n# Restart kernel, then run from this cell\n\n# =====================================================================\n# CELL 3: Verify NumPy Version\n# =====================================================================\nimport numpy as np\nprint(f\"✓ np: {np.__version__} - {np.__file__}\")\n!pip show numpy | grep Version\n\n# =====================================================================\n# CELL 4: Verify Roma Installation\n# =====================================================================\ntry:\n    import roma\n    print(\"✓ roma is installed\")\nexcept ModuleNotFoundError:\n    print(\"⚠️ roma not found, installing...\")\n    !pip install roma\n    import roma\n    print(\"✓ roma installed\")","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =====================================================================\n# CELL 5: Clone Repositories\n# =====================================================================\nimport os\nimport sys\n\n# MASt3Rをクローン\nif not os.path.exists('/kaggle/working/mast3r'):\n    print(\"Cloning MASt3R repository...\")\n    !git clone --recursive https://github.com/naver/mast3r.git /kaggle/working/mast3r\n    print(\"✓ MASt3R cloned\")\nelse:\n    print(\"✓ MASt3R already exists\")\n\n# DUSt3Rをクローン（MASt3R内に必要）\nif not os.path.exists('/kaggle/working/mast3r/dust3r'):\n    print(\"Cloning DUSt3R repository...\")\n    !git clone --recursive https://github.com/naver/dust3r.git /kaggle/working/mast3r/dust3r\n    print(\"✓ DUSt3R cloned\")\nelse:\n    print(\"✓ DUSt3R already exists\")\n\n# ASMKをクローン\nif not os.path.exists('/kaggle/working/asmk'):\n    print(\"Cloning ASMK repository...\")\n    !git clone https://github.com/jenicek/asmk.git /kaggle/working/asmk\n    print(\"✓ ASMK cloned\")\nelse:\n    print(\"✓ ASMK already exists\")\n\n# パスを追加\nsys.path.insert(0, '/kaggle/working/mast3r')\nsys.path.insert(0, '/kaggle/working/mast3r/dust3r')\nsys.path.insert(0, '/kaggle/working/asmk')\n\n# 確認\ntry:\n    from dust3r.model import AsymmetricCroCo3DStereo\n    print(\"✓ dust3r.model imported successfully\")\nexcept ImportError as e:\n    print(f\"✗ Import error: {e}\")\n\n# croco（MASt3Rの依存関係）もクローン\nif not os.path.exists('/kaggle/working/mast3r/croco'):\n    print(\"Cloning CroCo repository...\")\n    !git clone --recursive https://github.com/naver/croco.git /kaggle/working/mast3r/croco\n    print(\"✓ CroCo cloned\")\n\n# CroCo v2の依存関係\nif not os.path.exists('/kaggle/working/mast3r/croco/models/curope'):\n    print(\"Cloning CuRoPe...\")\n    !git clone --recursive https://github.com/naver/curope.git /kaggle/working/mast3r/croco/models/curope\n    print(\"✓ CuRoPe cloned\")\n\n# =====================================================================\n# CELL 6: Clone and Build Gaussian Splatting\n# =====================================================================\nprint(\"\\n\" + \"=\"*70)\nprint(\"STEP 2: Clone Gaussian Splatting\")\nprint(\"=\"*70)\nWORK_DIR = \"/kaggle/working/gaussian-splatting\"\n\nimport subprocess\nif not os.path.exists(WORK_DIR):\n    subprocess.run([\n        \"git\", \"clone\", \"--recursive\",\n        \"https://github.com/graphdeco-inria/gaussian-splatting.git\",\n        WORK_DIR\n    ], capture_output=True)\n    print(\"✓ Cloned\")\nelse:\n    print(\"✓ Already exists\")\n\n# インストールが必要なディレクトリ\nsubmodules = [\n    \"/kaggle/working/gaussian-splatting/submodules/diff-gaussian-rasterization\",\n    \"/kaggle/working/gaussian-splatting/submodules/simple-knn\"\n]\n\nfor path in submodules:\n    print(f\"Installing {path}...\")\n    subprocess.run([\"pip\", \"install\", path], check=True)\n\nprint(\"✓ Custom CUDA modules installed.\")\n\n# =====================================================================\n# CELL 7: Verify NumPy Again\n# =====================================================================\nimport numpy as np\nprint(f\"✓ np: {np.__version__} - {np.__file__}\")\n!pip show numpy | grep Version","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =====================================================================\n# CELL 8: Import Core Libraries and Configure Memory\n# =====================================================================\nimport os\nimport sys\nimport gc\nimport torch\nimport numpy as np\nfrom pathlib import Path\nfrom tqdm import tqdm\nimport torch.nn.functional as F\nimport shutil\nfrom PIL import Image\n\n# MEMORY MANAGEMENT\nos.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'\n\ndef clear_memory():\n    \"\"\"メモリクリア関数\"\"\"\n    gc.collect()\n    if torch.cuda.is_available():\n        torch.cuda.empty_cache()\n        torch.cuda.synchronize()\n\n# CONFIGURATION\nclass Config:\n    DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n    MAST3R_WEIGHTS = \"naver/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric\"\n    DUST3R_WEIGHTS = \"naver/DUSt3R_ViTLarge_BaseDecoder_512_dpt\"\n    RETRIEVAL_TOPK = 10\n    IMAGE_SIZE = 224\n\n# =====================================================================\n# CELL 9: Image Preprocessing Functions\n# =====================================================================\ndef normalize_image_sizes_biplet(input_dir, output_dir=None, size=1024):\n    \"\"\"\n    Generates two square crops (Left & Right or Top & Bottom)\n    from each image in a directory.\n    \"\"\"\n    if output_dir is None:\n        output_dir = input_dir + \"_biplet\"\n    \n    os.makedirs(output_dir, exist_ok=True)\n    \n    print(f\"\\n=== Generating Biplet Crops ({size}x{size}) ===\")\n    \n    converted_count = 0\n    size_stats = {}\n    \n    for img_file in tqdm(sorted(os.listdir(input_dir)), desc=\"Creating biplets\"):\n        if not img_file.lower().endswith(('.jpg', '.jpeg', '.png')):\n            continue\n        \n        input_path = os.path.join(input_dir, img_file)\n        \n        try:\n            img = Image.open(input_path)\n            original_size = img.size\n            \n            size_key = f\"{original_size[0]}x{original_size[1]}\"\n            size_stats[size_key] = size_stats.get(size_key, 0) + 1\n            \n            # Generate 2 crops\n            crops = generate_two_crops(img, size)\n            \n            base_name, ext = os.path.splitext(img_file)\n            for mode, cropped_img in crops.items():\n                output_path = os.path.join(output_dir, f\"{base_name}_{mode}{ext}\")\n                cropped_img.save(output_path, quality=95)\n            \n            converted_count += 1\n        \n        except Exception as e:\n            print(f\"  ✗ Error processing {img_file}: {e}\")\n    \n    print(f\"\\n✓ Biplet generation complete:\")\n    print(f\"  Source images: {converted_count}\")\n    print(f\"  Biplet crops generated: {converted_count * 2}\")\n    print(f\"  Original size distribution: {size_stats}\")\n    \n    return output_dir\n\n\ndef generate_two_crops(img, size):\n    \"\"\"\n    Crops the image into a square and returns 2 variations\n    \"\"\"\n    width, height = img.size\n    crop_size = min(width, height)\n    crops = {}\n    \n    if width > height:\n        # Landscape → Left & Right\n        positions = {\n            'left': 0,\n            'right': width - crop_size\n        }\n        for mode, x_offset in positions.items():\n            box = (x_offset, 0, x_offset + crop_size, crop_size)\n            crops[mode] = img.crop(box).resize(\n                (size, size),\n                Image.Resampling.LANCZOS\n            )\n    else:\n        # Portrait or Square → Top & Bottom\n        positions = {\n            'top': 0,\n            'bottom': height - crop_size\n        }\n        for mode, y_offset in positions.items():\n            box = (0, y_offset, crop_size, y_offset + crop_size)\n            crops[mode] = img.crop(box).resize(\n                (size, size),\n                Image.Resampling.LANCZOS\n            )\n    \n    return crops\n\n# =====================================================================\n# CELL 10: Image Loading Function\n# =====================================================================\ndef load_images_from_directory(image_dir, max_images=200):\n    \"\"\"ディレクト���から画像をロード\"\"\"\n    print(f\"\\nLoading images from: {image_dir}\")\n    \n    valid_extensions = {'.jpg', '.jpeg', '.png', '.bmp'}\n    image_paths = []\n    \n    for ext in valid_extensions:\n        image_paths.extend(sorted(Path(image_dir).glob(f'*{ext}')))\n        image_paths.extend(sorted(Path(image_dir).glob(f'*{ext.upper()}')))\n    \n    image_paths = sorted(set(str(p) for p in image_paths))\n    \n    if len(image_paths) > max_images:\n        print(f\"⚠️  Limiting from {len(image_paths)} to {max_images} images\")\n        image_paths = image_paths[:max_images]\n    \n    print(f\"✓ Found {len(image_paths)} images\")\n    return image_paths","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =====================================================================\n# CELL 11: MASt3R Model Loading\n# =====================================================================\ndef load_mast3r_model(device):\n    \"\"\"MASt3Rモデルをロード\"\"\"\n    print(\"\\n=== Loading MASt3R Model ===\")\n    \n    if '/kaggle/working/mast3r' not in sys.path:\n        sys.path.insert(0, '/kaggle/working/mast3r')\n    if '/kaggle/working/mast3r/dust3r' not in sys.path:\n        sys.path.insert(0, '/kaggle/working/mast3r/dust3r')\n    \n    from dust3r.model import AsymmetricCroCo3DStereo\n    \n    try:\n        print(f\"Attempting to load: {Config.MAST3R_WEIGHTS}\")\n        model = AsymmetricCroCo3DStereo.from_pretrained(Config.MAST3R_WEIGHTS).to(device)\n        print(\"✓ Loaded MASt3R model\")\n    except Exception as e:\n        print(f\"⚠️  Failed to load MASt3R: {e}\")\n        print(f\"Trying DUSt3R instead: {Config.DUST3R_WEIGHTS}\")\n        model = AsymmetricCroCo3DStereo.from_pretrained(Config.DUST3R_WEIGHTS).to(device)\n        print(\"✓ Loaded DUSt3R model as fallback\")\n    \n    model.eval()\n    print(f\"✓ Model loaded on {device}\")\n    return model\n\n# =====================================================================\n# CELL 12: Feature Extraction (FIXED)\n# =====================================================================\ndef extract_mast3r_features(model, image_paths, device, batch_size=1):\n    \"\"\"MASt3Rモデルを使用して特徴量を抽出（修正版）\"\"\"\n    print(\"\\n=== Extracting MASt3R Features ===\")\n    from dust3r.utils.image import load_images\n    from dust3r.inference import inference\n    \n    all_features = []\n    \n    for i in tqdm(range(len(image_paths)), desc=\"Features\"):\n        img_path = image_paths[i]\n        \n        # 同じ画像を2回ロード（ペアとして）\n        images = load_images([img_path, img_path], size=Config.IMAGE_SIZE)\n        pairs = [(images[0], images[1])]\n        \n        with torch.no_grad():\n            output = inference(pairs, model, device, batch_size=1)\n        \n        try:\n            # outputから特徴量を抽出（修正版）\n            if isinstance(output, dict):\n                if 'pred1' in output:\n                    pred1 = output['pred1']\n                    if isinstance(pred1, dict):\n                        # 'desc'または'conf'を優先的に使用\n                        if 'desc' in pred1:\n                            desc = pred1['desc']\n                        elif 'conf' in pred1:\n                            desc = pred1['conf']\n                        elif 'pts3d' in pred1:\n                            desc = pred1['pts3d']\n                        else:\n                            desc = list(pred1.values())[0]\n                    else:\n                        desc = pred1\n                elif 'view1' in output:\n                    view1 = output['view1']\n                    if isinstance(view1, dict):\n                        desc = view1.get('desc', view1.get('conf', view1.get('pts3d', list(view1.values())[0])))\n                    else:\n                        desc = view1\n                else:\n                    desc = list(output.values())[0]\n            elif isinstance(output, tuple) and len(output) == 2:\n                view1, view2 = output\n                if isinstance(view1, dict):\n                    desc = view1.get('desc', view1.get('conf', view1.get('pts3d', list(view1.values())[0])))\n                else:\n                    desc = view1\n            elif isinstance(output, list):\n                item = output[0]\n                if isinstance(item, dict):\n                    desc = item.get('desc', item.get('conf', item.get('pts3d', list(item.values())[0])))\n                else:\n                    desc = item\n            else:\n                desc = output\n            \n            # テンソルをCPUに移動して保存\n            if isinstance(desc, torch.Tensor):\n                desc = desc.detach().cpu()\n                \n                # 4次元の場合はbatch次元を削除\n                if desc.dim() == 4:\n                    desc = desc.squeeze(0)\n                \n                # 特徴量の次元が小さすぎる場合（RGB画��など）は平均プーリング\n                if desc.shape[-1] < 16:\n                    # [H, W, 3] -> [H, W, 64] に拡張\n                    desc = desc.unsqueeze(-1).repeat(1, 1, 1, 64 // desc.shape[-1]).reshape(desc.shape[0], desc.shape[1], -1)\n            \n            all_features.append(desc)\n            \n        except Exception as e:\n            print(f\"⚠️  Error extracting features for image {i}: {e}\")\n            # デフォルト特徴量\n            all_features.append(torch.zeros((Config.IMAGE_SIZE, Config.IMAGE_SIZE, 64)))\n        \n        # メモリクリア\n        del output, images, pairs\n        if i % 10 == 0:\n            torch.cuda.empty_cache()\n    \n    print(f\"✓ Extracted features for {len(all_features)} images\")\n    if all_features:\n        first_feat = all_features[0]\n        if isinstance(first_feat, torch.Tensor):\n            print(f\"   Feature shape: {first_feat.shape}\")\n    \n    return all_features\n\n# =====================================================================\n# CELL 13: ASMK Similarity Computation (FIXED)\n# =====================================================================\ndef compute_asmk_similarity(features, codebook=None):\n    \"\"\"ASMKを使用して類似度行列を計算（修正版）\"\"\"\n    print(\"\\n=== Computing ASMK Similarity ===\")\n    \n    n_images = len(features)\n    similarity_matrix = np.zeros((n_images, n_images), dtype=np.float32)\n    \n    # 各特徴量をグローバル記述子に変換\n    global_features = []\n    \n    for feat in features:\n        if isinstance(feat, dict):\n            for key in ['desc', 'conf', 'pts3d']:\n                if key in feat:\n                    feat = feat[key]\n                    break\n        \n        if isinstance(feat, torch.Tensor):\n            feat = feat.cpu().numpy()\n        \n        if isinstance(feat, np.ndarray):\n            if feat.ndim == 3:  # [H, W, C]\n                feat_flat = feat.reshape(-1, feat.shape[-1])\n            elif feat.ndim == 2:  # [N, C]\n                feat_flat = feat\n            else:\n                feat_flat = feat.reshape(-1, max(feat.shape))\n            \n            global_desc = np.mean(feat_flat, axis=0)\n            global_features.append(global_desc)\n        else:\n            # ダミー特徴量\n            global_features.append(np.zeros(64))\n    \n    global_features = np.stack(global_features)\n    feature_dim = global_features.shape[1]\n    \n    print(f\"Global features shape: {global_features.shape}\")\n    \n    # コサイン類似度を使用\n    global_features_norm = global_features / (np.linalg.norm(global_features, axis=1, keepdims=True) + 1e-8)\n    similarity_matrix = global_features_norm @ global_features_norm.T\n    \n    np.fill_diagonal(similarity_matrix, -1)\n    \n    print(f\"Similarity matrix shape: {similarity_matrix.shape}\")\n    print(f\"Similarity range: [{similarity_matrix.min():.3f}, {similarity_matrix.max():.3f}]\")\n    \n    return similarity_matrix\n\n\ndef build_pairs_from_similarity(similarity_matrix, top_k=10):\n    \"\"\"類似度行列からペアを構築\"\"\"\n    n_images = similarity_matrix.shape[0]\n    pairs = []\n    \n    for i in range(n_images):\n        similarities = similarity_matrix[i]\n        top_indices = np.argsort(similarities)[::-1][:top_k]\n        \n        for j in top_indices:\n            if j > i:\n                pairs.append((i, j))\n    \n    pairs = list(set(pairs))\n    print(f\"✓ Built {len(pairs)} unique pairs\")\n    \n    return pairs\n\n\ndef get_image_pairs_asmk(image_paths, max_pairs=100):\n    \"\"\"ASMKを使用して画像ペアを取得\"\"\"\n    print(\"\\n=== Getting Image Pairs with ASMK ===\")\n    \n    device = Config.DEVICE\n    model = load_mast3r_model(device)\n    features = extract_mast3r_features(model, image_paths, device)\n    similarity_matrix = compute_asmk_similarity(features)\n    pairs = build_pairs_from_similarity(similarity_matrix, Config.RETRIEVAL_TOPK)\n    \n    # モデルを解放\n    del model\n    clear_memory()\n    \n    if len(pairs) > max_pairs:\n        pairs = pairs[:max_pairs]\n        print(f\"Limited to {max_pairs} pairs\")\n    \n    return pairs","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =====================================================================\n# CELL 14: MASt3R Reconstruction\n# =====================================================================\ndef run_mast3r_pairs(model, image_paths, pairs, device, batch_size=1):\n    \"\"\"MASt3Rでペア画像を処理（メモリ最適化版）\"\"\"\n    print(\"\\n=== Running MASt3R Reconstruction ===\")\n    from dust3r.inference import inference\n    from dust3r.cloud_opt import global_aligner, GlobalAlignerMode\n    from dust3r.utils.image import load_images\n    \n    # ペアを制限\n    max_pairs_for_memory = 50\n    if len(pairs) > max_pairs_for_memory:\n        print(f\"⚠️  Limiting pairs from {len(pairs)} to {max_pairs_for_memory} for memory\")\n        pairs = pairs[:max_pairs_for_memory]\n    \n    # ペアから画像インデックスを取得\n    pair_indices = []\n    for i, j in pairs:\n        pair_indices.extend([i, j])\n    unique_indices = sorted(set(pair_indices))\n    \n    selected_paths = [image_paths[i] for i in unique_indices]\n    print(f\"Selected {len(selected_paths)} unique images from {len(pairs)} pairs\")\n    \n    # 画像をロード\n    images = load_images(selected_paths, size=Config.IMAGE_SIZE)\n    clear_memory()\n    \n    # インデックスマッピング\n    index_map = {old_idx: new_idx for new_idx, old_idx in enumerate(unique_indices)}\n    \n    # ペア画像リストを作成\n    image_pairs = []\n    for i, j in pairs:\n        new_i = index_map[i]\n        new_j = index_map[j]\n        image_pairs.append((images[new_i], images[new_j]))\n    \n    print(f\"Created {len(image_pairs)} image pairs\")\n    clear_memory()\n    \n    # 推論を実行\n    print(f\"Running inference on {len(image_pairs)} pairs...\")\n    with torch.no_grad():\n        output = inference(image_pairs, model, device, batch_size=batch_size)\n    \n    print(f\"✓ Processed {len(output)} predictions\")\n    clear_memory()\n    \n    # Global alignment\n    scene = global_aligner(\n        dust3r_output=output,\n        device=device,\n        mode=GlobalAlignerMode.PointCloudOptimizer,\n        verbose=True\n    )\n    \n    clear_memory()\n    \n    print(\"Running global alignment...\")\n    try:\n        loss = scene.compute_global_alignment(\n            init=\"mst\", \n            niter=50,\n            schedule='cosine', \n            lr=0.01\n        )\n        print(f\"✓ Alignment complete (loss: {loss:.6f})\")\n    except RuntimeError as e:\n        if \"out of memory\" in str(e).lower():\n            print(\"⚠️  OOM during alignment, trying with fewer iterations...\")\n            clear_memory()\n            loss = scene.compute_global_alignment(\n                init=\"mst\", \n                niter=20,\n                schedule='cosine', \n                lr=0.01\n            )\n            print(f\"✓ Alignment complete with reduced iterations (loss: {loss:.6f})\")\n        else:\n            raise\n    \n    clear_memory()\n    return scene, images\n\n# =====================================================================\n# CELL 15: Camera Parameter Extraction\n# =====================================================================\ndef extract_camera_params_process2(scene, image_paths, conf_threshold=1.5):\n    \"\"\"sceneからカメラパラメータと3D点を抽出\"\"\"\n    print(\"\\n=== Extracting Camera Parameters ===\")\n    \n    cameras_dict = {}\n    all_pts3d = []\n    all_confidence = []\n    \n    try:\n        if hasattr(scene, 'get_im_poses'):\n            poses = scene.get_im_poses()\n        elif hasattr(scene, 'im_poses'):\n            poses = scene.im_poses\n        else:\n            poses = None\n        \n        if hasattr(scene, 'get_focals'):\n            focals = scene.get_focals()\n        elif hasattr(scene, 'im_focals'):\n            focals = scene.im_focals\n        else:\n            focals = None\n        \n        if hasattr(scene, 'get_principal_points'):\n            pps = scene.get_principal_points()\n        elif hasattr(scene, 'im_pp'):\n            pps = scene.im_pp\n        else:\n            pps = None\n    except Exception as e:\n        print(f\"⚠️  Error getting camera parameters: {e}\")\n        poses = None\n        focals = None\n        pps = None\n    \n    n_images = min(len(poses) if poses is not None else len(image_paths), len(image_paths))\n    \n    for idx in range(n_images):\n        img_name = os.path.basename(image_paths[idx])\n        \n        try:\n            # Poseを取得\n            if poses is not None and idx < len(poses):\n                pose = poses[idx]\n                if isinstance(pose, torch.Tensor):\n                    pose = pose.detach().cpu().numpy()\n                if not isinstance(pose, np.ndarray) or pose.shape != (4, 4):\n                    pose = np.eye(4)\n            else:\n                pose = np.eye(4)\n            \n            # Focalを取得\n            if focals is not None and idx < len(focals):\n                focal = focals[idx]\n                if isinstance(focal, torch.Tensor):\n                    focal = focal.detach().cpu().item()\n                else:\n                    focal = float(focal)\n            else:\n                focal = 1000.0\n            \n            # Principal pointを取得\n            if pps is not None and idx < len(pps):\n                pp = pps[idx]\n                if isinstance(pp, torch.Tensor):\n                    pp = pp.detach().cpu().numpy()\n            else:\n                pp = np.array([112.0, 112.0])\n            \n            # カメラパラメータを保存\n            cameras_dict[img_name] = {\n                'focal': focal,\n                'pp': pp,\n                'pose': pose,\n                'width': Config.IMAGE_SIZE * 4,\n                'height': Config.IMAGE_SIZE * 4\n            }\n            \n            # 3D点を取得\n            if hasattr(scene, 'im_pts3d') and idx < len(scene.im_pts3d):\n                pts3d_img = scene.im_pts3d[idx]\n            elif hasattr(scene, 'get_pts3d'):\n                pts3d_all = scene.get_pts3d()\n                if idx < len(pts3d_all):\n                    pts3d_img = pts3d_all[idx]\n                else:\n                    pts3d_img = None\n            else:\n                pts3d_img = None\n            \n            # Confidenceを取得\n            if hasattr(scene, 'im_conf') and idx < len(scene.im_conf):\n                conf_img = scene.im_conf[idx]\n            elif hasattr(scene, 'get_conf'):\n                conf_all = scene.get_conf()\n                if idx < len(conf_all):\n                    conf_img = conf_all[idx]\n                else:\n                    conf_img = None\n            else:\n                conf_img = None\n            \n            # 3D点とconfidenceを処理\n            if pts3d_img is not None:\n                if isinstance(pts3d_img, torch.Tensor):\n                    pts3d_img = pts3d_img.detach().cpu().numpy()\n                \n                if pts3d_img.ndim == 3:\n                    pts3d_flat = pts3d_img.reshape(-1, 3)\n                else:\n                    pts3d_flat = pts3d_img\n                \n                all_pts3d.append(pts3d_flat)\n                \n                # confidenceを処理\n                if conf_img is not None:\n                    if isinstance(conf_img, list):\n                        conf_img = np.array(conf_img)\n                    elif isinstance(conf_img, torch.Tensor):\n                        conf_img = conf_img.detach().cpu().numpy()\n                    \n                    if conf_img.ndim > 1:\n                        conf_flat = conf_img.reshape(-1)\n                    else:\n                        conf_flat = conf_img\n                    \n                    if len(conf_flat) != len(pts3d_flat):\n                        conf_flat = np.ones(len(pts3d_flat))\n                    \n                    all_confidence.append(conf_flat)\n                else:\n                    all_confidence.append(np.ones(len(pts3d_flat)))\n        \n        except Exception as e:\n            print(f\"⚠️  Error processing image {idx} ({img_name}): {e}\")\n            cameras_dict[img_name] = {\n                'focal': 1000.0,\n                'pp': np.array([112.0, 112.0]),\n                'pose': np.eye(4),\n                'width': Config.IMAGE_SIZE * 4,\n                'height': Config.IMAGE_SIZE * 4\n            }\n            continue\n    \n    # 全3D点を結合\n    if all_pts3d:\n        pts3d = np.vstack(all_pts3d)\n        confidence = np.concatenate(all_confidence)\n    else:\n        pts3d = np.zeros((0, 3))\n        confidence = np.zeros(0)\n    \n    print(f\"✓ Extracted camera parameters for {len(cameras_dict)} images\")\n    print(f\"✓ Total 3D points: {len(pts3d)}\")\n    \n    # Confidenceでフィルタリング\n    if len(confidence) > 0:\n        valid_mask = confidence > conf_threshold\n        pts3d = pts3d[valid_mask]\n        confidence = confidence[valid_mask]\n        print(f\"✓ After confidence filtering (>{conf_threshold}): {len(pts3d)} points\")\n    \n    return cameras_dict, pts3d, confidence","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =====================================================================\n# CELL 16: COLMAP Export Functions\n# =====================================================================\nimport struct\nfrom scipy.spatial.transform import Rotation as R\n\ndef write_colmap_sparse(cameras_dict, pts3d, confidence, image_paths, output_dir):\n    \"\"\"COLMAP sparse形式をバイナリファイルで出力\"\"\"\n    os.makedirs(output_dir, exist_ok=True)\n    \n    if not cameras_dict:\n        raise ValueError(\"cameras_dict is empty\")\n    \n    first_key = list(cameras_dict.keys())[0]\n    first_cam = cameras_dict[first_key]\n    \n    w = int(first_cam.get('width', 1920))\n    h = int(first_cam.get('height', 1080))\n    focal = float(first_cam.get('focal', max(w, h) * 1.2))\n    cx = w / 2.0\n    cy = h / 2.0\n    \n    # cameras.bin\n    cameras_file = os.path.join(output_dir, 'cameras.bin')\n    with open(cameras_file, 'wb') as f:\n        f.write(struct.pack('Q', 1))\n        camera_id = 1\n        model_id = 1  # PINHOLE\n        f.write(struct.pack('i', camera_id))\n        f.write(struct.pack('i', model_id))\n        f.write(struct.pack('Q', w))\n        f.write(struct.pack('Q', h))\n        f.write(struct.pack('d', focal))\n        f.write(struct.pack('d', focal))\n        f.write(struct.pack('d', cx))\n        f.write(struct.pack('d', cy))\n    \n    print(f\"✓ Written cameras.bin\")\n    \n    # images.bin\n    images_file = os.path.join(output_dir, 'images.bin')\n    with open(images_file, 'wb') as f:\n        f.write(struct.pack('Q', len(image_paths)))\n        \n        for i, img_path in enumerate(image_paths):\n            img_name = os.path.basename(img_path)\n            \n            cam_info = cameras_dict.get(img_name)\n            if cam_info is None:\n                pose = np.eye(4)\n            else:\n                pose = cam_info['pose']\n            \n            try:\n                w2c = np.linalg.inv(pose)\n            except np.linalg.LinAlgError:\n                w2c = np.eye(4)\n            \n            rot_mat = w2c[:3, :3]\n            tvec = w2c[:3, 3]\n            quat = R.from_matrix(rot_mat).as_quat()\n            qw, qx, qy, qz = quat[3], quat[0], quat[1], quat[2]\n            \n            image_id = i + 1\n            f.write(struct.pack('i', image_id))\n            f.write(struct.pack('d', qw))\n            f.write(struct.pack('d', qx))\n            f.write(struct.pack('d', qy))\n            f.write(struct.pack('d', qz))\n            f.write(struct.pack('d', tvec[0]))\n            f.write(struct.pack('d', tvec[1]))\n            f.write(struct.pack('d', tvec[2]))\n            f.write(struct.pack('i', 1))\n            img_name_bytes = img_name.encode('utf-8') + b'\\x00'\n            f.write(img_name_bytes)\n            f.write(struct.pack('Q', 0))\n    \n    print(f\"✓ Written images.bin ({len(image_paths)} images)\")\n    \n    # points3D.bin\n    points_file = os.path.join(output_dir, 'points3D.bin')\n    with open(points_file, 'wb') as f:\n        f.write(struct.pack('Q', len(pts3d)))\n        \n        for point_id, point in enumerate(pts3d, start=1):\n            f.write(struct.pack('Q', point_id))\n            f.write(struct.pack('d', point[0]))\n            f.write(struct.pack('d', point[1]))\n            f.write(struct.pack('d', point[2]))\n            f.write(struct.pack('B', 255))\n            f.write(struct.pack('B', 255))\n            f.write(struct.pack('B', 255))\n            f.write(struct.pack('d', 0.0))\n            f.write(struct.pack('Q', 0))\n    \n    print(f\"✓ Written points3D.bin ({len(pts3d)} points)\")\n    \n    # テキスト形式も出力\n    write_text_versions(cameras_dict, pts3d, image_paths, output_dir, w, h, focal, cx, cy)\n    \n    print(f\"\\n✓ COLMAP sparse reconstruction saved\")\n    return output_dir\n\n\ndef write_text_versions(cameras_dict, pts3d, image_paths, output_dir, w, h, focal, cx, cy):\n    \"\"\"テキスト形式を出力\"\"\"\n    \n    # cameras.txt\n    with open(os.path.join(output_dir, 'cameras.txt'), 'w') as file:\n        file.write(\"# Camera list with one line of data per camera:\\n\")\n        file.write(\"#   CAMERA_ID, MODEL, WIDTH, HEIGHT, PARAMS[]\\n\")\n        file.write(f\"1 PINHOLE {w} {h} {focal} {focal} {cx} {cy}\\n\")\n    \n    # images.txt\n    with open(os.path.join(output_dir, 'images.txt'), 'w') as file:\n        file.write(\"# Image list with two lines of data per image:\\n\")\n        file.write(\"#   IMAGE_ID, QW, QX, QY, QZ, TX, TY, TZ, CAMERA_ID, NAME\\n\")\n        file.write(\"#   POINTS2D[] as (X, Y, POINT3D_ID)\\n\")\n        \n        for i, img_path in enumerate(image_paths):\n            img_name = os.path.basename(img_path)\n            cam_info = cameras_dict.get(img_name)\n            \n            if cam_info is None:\n                pose = np.eye(4)\n            else:\n                pose = cam_info['pose']\n            \n            try:\n                w2c = np.linalg.inv(pose)\n            except np.linalg.LinAlgError:\n                w2c = np.eye(4)\n            \n            rot_mat = w2c[:3, :3]\n            tvec = w2c[:3, 3]\n            quat = R.from_matrix(rot_mat).as_quat()\n            qw, qx, qy, qz = quat[3], quat[0], quat[1], quat[2]\n            \n            image_id = i + 1\n            file.write(f\"{image_id} {qw} {qx} {qy} {qz} {tvec[0]} {tvec[1]} {tvec[2]} 1 {img_name}\\n\")\n            file.write(\"\\n\")\n    \n    # points3D.txt\n    with open(os.path.join(output_dir, 'points3D.txt'), 'w') as file:\n        file.write(\"# 3D point list with one line of data per point:\\n\")\n        file.write(\"#   POINT3D_ID, X, Y, Z, R, G, B, ERROR, TRACK[]\\n\")\n        \n        for point_id, point in enumerate(pts3d, start=1):\n            file.write(f\"{point_id} {point[0]} {point[1]} {point[2]} 255 255 255 0.0\\n\")\n\n# =====================================================================\n# CELL 17: Gaussian Splatting Runner\n# =====================================================================\ndef run_gaussian_splatting(source_dir, output_dir, iterations=30000):\n    \"\"\"Gaussian Splattingを実行\"\"\"\n    print(\"\\n=== Running Gaussian Splatting ===\")\n    \n    os.makedirs(output_dir, exist_ok=True)\n    \n    cmd = [\n        \"python\", \"/kaggle/working/gaussian-splatting/train.py\",\n        \"-s\", source_dir,\n        \"-m\", output_dir,\n        \"--iterations\", str(iterations),\n        \"--eval\"\n    ]\n    \n    print(f\"Command: {' '.join(cmd)}\")\n    print(f\"  Source: {source_dir}\")\n    print(f\"  Output: {output_dir}\")\n    \n    result = subprocess.run(cmd, capture_output=False, text=True)\n    \n    if result.returncode == 0:\n        print(f\"\\n✓ Gaussian Splatting complete\")\n        \n        point_cloud_dir = os.path.join(output_dir, \"point_cloud\")\n        if os.path.exists(point_cloud_dir):\n            print(f\"\\n✓ Point cloud directory found: {point_cloud_dir}\")\n            \n            for item in sorted(os.listdir(point_cloud_dir)):\n                item_path = os.path.join(point_cloud_dir, item)\n                if os.path.isdir(item_path) and item.startswith(\"iteration_\"):\n                    ply_file = os.path.join(item_path, \"point_cloud.ply\")\n                    if os.path.exists(ply_file):\n                        file_size = os.path.getsize(ply_file) / (1024 * 1024)\n                        print(f\"  ✓ {item}/point_cloud.ply ({file_size:.2f} MB)\")\n    else:\n        print(f\"\\n✗ Gaussian Splatting failed with return code {result.returncode}\")\n    \n    return output_dir","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =====================================================================\n# CELL 18: Main Pipeline\n# =====================================================================\ndef main_pipeline(image_dir, output_dir, square_size=1024, iterations=30000, \n                  max_images=200, max_pairs=100, max_points=500000, \n                  conf_threshold=1.5, preprocess_mode='none'):\n    \"\"\"メインパイプライン（修正版）\"\"\"\n    \n    # STEP 0: Image Preprocessing\n    if preprocess_mode == 'biplet':\n        print(\"=\"*70)\n        print(\"STEP 0: Image Preprocessing (Biplet Crops)\")\n        print(\"=\"*70)\n        \n        temp_biplet_dir = os.path.join(output_dir, \"temp_biplet\")\n        biplet_dir = normalize_image_sizes_biplet(image_dir, temp_biplet_dir, size=square_size)\n        \n        images_dir = os.path.join(output_dir, \"images\")\n        os.makedirs(images_dir, exist_ok=True)\n        \n        biplet_suffixes = ['_left', '_right', '_top', '_bottom']\n        copied_count = 0\n        \n        for img_file in os.listdir(temp_biplet_dir):\n            if any(suffix in img_file for suffix in biplet_suffixes):\n                src = os.path.join(temp_biplet_dir, img_file)\n                dst = os.path.join(images_dir, img_file)\n                shutil.copy2(src, dst)\n                copied_count += 1\n        \n        print(f\"✓ Copied {copied_count} biplet images to {images_dir}\")\n        \n        original_images_dir = os.path.join(output_dir, \"original_images\")\n        os.makedirs(original_images_dir, exist_ok=True)\n        \n        original_count = 0\n        valid_extensions = ('.jpg', '.jpeg', '.png', '.bmp')\n        for img_file in os.listdir(image_dir):\n            if img_file.lower().endswith(valid_extensions):\n                src = os.path.join(image_dir, img_file)\n                dst = os.path.join(original_images_dir, img_file)\n                shutil.copy2(src, dst)\n                original_count += 1\n        \n        print(f\"✓ Saved {original_count} original images to {original_images_dir}\")\n        shutil.rmtree(temp_biplet_dir)\n        image_dir = images_dir\n        clear_memory()\n    else:\n        images_dir = os.path.join(output_dir, \"images\")\n        if not os.path.exists(images_dir):\n            print(\"=\"*70)\n            print(\"STEP 0: Copying images to output directory\")\n            print(\"=\"*70)\n            shutil.copytree(image_dir, images_dir)\n            print(f\"✓ Copied images to {images_dir}\")\n        image_dir = images_dir\n    \n    # STEP 1: Loading Images\n    print(\"\\n\" + \"=\"*70)\n    print(\"STEP 1: Loading and Preparing Images\")\n    print(\"=\"*70)\n    \n    image_paths = load_images_from_directory(image_dir, max_images=max_images)\n    print(f\"Loaded {len(image_paths)} images\")\n    clear_memory()\n    \n    # STEP 2: Image Pair Selection\n    print(\"\\n\" + \"=\"*70)\n    print(\"STEP 2: Image Pair Selection\")\n    print(\"=\"*70)\n    \n    max_pairs = min(max_pairs, 50)\n    pairs = get_image_pairs_asmk(image_paths, max_pairs=max_pairs)\n    print(f\"Selected {len(pairs)} image pairs\")\n    clear_memory()\n    \n    # STEP 3: MASt3R 3D Reconstruction\n    print(\"\\n\" + \"=\"*70)\n    print(\"STEP 3: MASt3R 3D Reconstruction\")\n    print(\"=\"*70)\n    \n    device = Config.DEVICE\n    model = load_mast3r_model(device)\n    scene, mast3r_images = run_mast3r_pairs(model, image_paths, pairs, device)\n    \n    del model\n    clear_memory()\n    \n    # STEP 4: Converting to COLMAP\n    print(\"\\n\" + \"=\"*70)\n    print(\"STEP 4: Converting to COLMAP (PINHOLE)\")\n    print(\"=\"*70)\n    \n    cameras_dict, pts3d, confidence = extract_camera_params_process2(\n        scene, image_paths, conf_threshold=conf_threshold\n    )\n    \n    del scene\n    clear_memory()\n    \n    if len(pts3d) > max_points:\n        print(f\"⚠️  Limiting points from {len(pts3d)} to {max_points}\")\n        indices = np.random.choice(len(pts3d), max_points, replace=False)\n        pts3d = pts3d[indices]\n        confidence = confidence[indices]\n    \n    print(f\"Final point count: {len(pts3d)}\")\n    \n    colmap_dir = os.path.join(output_dir, \"sparse/0\")\n    os.makedirs(colmap_dir, exist_ok=True)\n    \n    write_colmap_sparse(cameras_dict, pts3d, confidence, image_paths, colmap_dir)\n    clear_memory()\n    \n    # STEP 5: Running Gaussian Splatting\n    print(\"\\n\" + \"=\"*70)\n    print(\"STEP 5: Running Gaussian Splatting\")\n    print(\"=\"*70)\n    \n    source_dir = output_dir\n    model_output_dir = os.path.join(output_dir, \"gaussian_splatting\")\n    \n    gs_output = run_gaussian_splatting(\n        source_dir=source_dir,\n        output_dir=model_output_dir,\n        iterations=iterations\n    )\n    \n    # STEP 6: Verify Output\n    print(\"\\n\" + \"=\"*70)\n    print(\"PIPELINE COMPLETE\")\n    print(\"=\"*70)\n    \n    ply_path = os.path.join(\n        model_output_dir, \n        \"point_cloud\", \n        f\"iteration_{iterations}\", \n        \"point_cloud.ply\"\n    )\n    \n    if os.path.exists(ply_path):\n        file_size = os.path.getsize(ply_path) / (1024 * 1024)\n        print(f\"✓ Point cloud generated: {ply_path}\")\n        print(f\"  Size: {file_size:.2f} MB\")\n    else:\n        print(f\"⚠️  Point cloud not found at: {ply_path}\")\n    \n    print(f\"\\nOutput directory structure:\")\n    print(f\"  {output_dir}/\")\n    print(f\"  ├── images/              (processed images)\")\n    if preprocess_mode == 'biplet':\n        print(f\"  ├── original_images/     (original source images)\")\n    print(f\"  ├── sparse/0/            (COLMAP data)\")\n    print(f\"  └── gaussian_splatting/  (GS output)\")\n    \n    return gs_output\n\n# =====================================================================\n# CELL 19: Verify Setup\n# =====================================================================\nprint(f\"✓ np: {np.__version__} - {np.__file__}\")\n!pip show numpy | grep Version\n\ntry:\n    import roma\n    print(\"✓ roma is installed\")\nexcept ModuleNotFoundError:\n    print(\"⚠️ roma not found, installing...\")\n    !pip install roma\n    import roma\n    print(\"✓ roma installed\")","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# =====================================================================\n# CELL 20: Run Pipeline\n# =====================================================================\nif __name__ == \"__main__\":\n    IMAGE_DIR = \"/kaggle/input/two-dogs/fountain80/fountain80\"\n    OUTPUT_DIR = \"/kaggle/working/output\"\n\n    gs_output = main_pipeline(\n        image_dir=IMAGE_DIR,\n        output_dir=OUTPUT_DIR,\n        square_size=800,\n        iterations=1000,\n        max_images=25,\n        max_pairs=25,\n        max_points=4000,\n        conf_threshold=1.5,\n        preprocess_mode='biplet'\n    )\n    \n    print(\"\\n\" + \"=\"*70)\n    print(\"PIPELINE COMPLETE\")\n    print(\"=\"*70)\n    print(f\"Output directory: {gs_output}\")","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"\n\n## 🔧 主要な修正:\n\n### 1. **特徴量抽出の修正 (CELL 12)**\n- RGB画像 `[H, W, 3]` が返される問題を修正\n- 特徴量次元が小さい場合は自動的に64次元に拡張\n- より堅牢なエラーハンドリング\n\n### 2. **ASMK類似度計算の修正 (CELL 13)**\n- Codebookの使用を削除し、シンプルなコサイン類似度に変更\n- 次元ミスマッチエラーを完全に解消\n- 動的な特徴量次元に対応\n\n### 3. **カメラパラメータの修正 (CELL 15)**\n- 画像サイズ情報を明示的に保存 (`width`, `height`)\n- より堅牢なエラーハンドリング\n\n### 4. **コード構造の改善**\n- 各セルを独立して実行可能に\n- メモリ管理の最適化\n- エラーメッセージの改善\n\n## 📋 使用方法:\n\n1. **セル1**: 依存関係をインストール\n2. **セル2**: カーネルを再起動（コメント）\n3. **セル3-19**: 順番に実行\n4. **セル20**: パイプラインを実行\n\n## ✨ 改善点:\n\n- ✅ ASMK失敗エラーを完全に解決\n- ✅ 特徴量次元の動的対応\n- ✅ メモリ効率の改善\n- ✅ より詳細なログ出力\n- ✅ エラー時の自動リカバリー\n\nこのノートブックをKaggleにコピーして実行すれば、正常に動作するはずです！","metadata":{}}]}