Hugging Face's logo

Datasets:
stpete2
/
ipynb

License:
Dataset card Files
xet
Community
stpete2 commited on
Commit
6beb136
·
verified ·
1 Parent(s): 1e74673

Delete biplet_asmk_mast3r_ps2_gs_kg_32_colab_01.ipynb

Browse files
Files changed (1) hide show
  1. biplet_asmk_mast3r_ps2_gs_kg_32_colab_01.ipynb +0 -1365
biplet_asmk_mast3r_ps2_gs_kg_32_colab_01.ipynb DELETED
@@ -1,1365 +0,0 @@
1
- {
2
- "metadata": {
3
- "kernelspec": {
4
- "name": "python3",
5
- "display_name": "Python 3",
6
- "language": "python"
7
- },
8
- "language_info": {
9
- "name": "python",
10
- "version": "3.12.12",
11
- "mimetype": "text/x-python",
12
- "codemirror_mode": {
13
- "name": "ipython",
14
- "version": 3
15
- },
16
- "pygments_lexer": "ipython3",
17
- "nbconvert_exporter": "python",
18
- "file_extension": ".py"
19
- },
20
- "colab": {
21
- "provenance": [],
22
- "gpuType": "T4"
23
- },
24
- "accelerator": "GPU",
25
- "kaggle": {
26
- "accelerator": "nvidiaTeslaT4",
27
- "dataSources": [
28
- {
29
- "sourceId": 14571475,
30
- "sourceType": "datasetVersion",
31
- "datasetId": 1429416
32
- }
33
- ],
34
- "dockerImageVersionId": 31260,
35
- "isInternetEnabled": true,
36
- "language": "python",
37
- "sourceType": "notebook",
38
- "isGpuEnabled": true
39
- }
40
- },
41
- "nbformat_minor": 0,
42
- "nbformat": 4,
43
- "cells": [
44
- {
45
- "cell_type": "markdown",
46
- "source": [
47
- "# **biplet-asmk-mast3r-ps2-gs-kg-32-colab**\n",
48
- "\n"
49
- ],
50
- "metadata": {
51
- "id": "qDQLX3PArmh8"
52
- }
53
- },
54
- {
55
- "cell_type": "markdown",
56
- "source": [
57
- "https://huggingface.co/datasets/stpete2/ipynb/blob/main/biplet-asmk-mast3r-ps2-gs-kg-32.ipynb"
58
- ],
59
- "metadata": {
60
- "id": "Yhla_oBUjLmD"
61
- }
62
- },
63
- {
64
- "cell_type": "code",
65
- "source": [
66
- "#これを元にcolab化 2025/01/22 16:00"
67
- ],
68
- "metadata": {
69
- "id": "UyF0gaG8jOXu"
70
- },
71
- "execution_count": null,
72
- "outputs": []
73
- },
74
- {
75
- "cell_type": "markdown",
76
- "source": [
77
- "v.32 全面見直し"
78
- ],
79
- "metadata": {
80
- "id": "uNZNREeejLmD"
81
- }
82
- },
83
- {
84
- "cell_type": "code",
85
- "source": [],
86
- "metadata": {
87
- "trusted": true,
88
- "id": "yH63Q7yCjLmE"
89
- },
90
- "outputs": [],
91
- "execution_count": null
92
- },
93
- {
94
- "cell_type": "code",
95
- "source": [
96
- "# =====================================================================\n",
97
- "# CELL 1: Install Dependencies\n",
98
- "# =====================================================================\n",
99
- "!pip install roma einops timm huggingface_hub\n",
100
- "!pip install opencv-python pillow tqdm pyaml cython plyfile\n",
101
- "!pip install pycolmap trimesh\n",
102
- "!pip uninstall -y numpy scipy\n",
103
- "!pip install numpy==1.26.4 scipy==1.11.4\n",
104
- "break"
105
- ],
106
- "metadata": {
107
- "trusted": true,
108
- "id": "h5Exo6FBjLmE"
109
- },
110
- "outputs": [],
111
- "execution_count": null
112
- },
113
- {
114
- "cell_type": "code",
115
- "source": [
116
- "# =====================================================================\n",
117
- "# CELL 2: Restart Kernel (Run this after Cell 1)\n",
118
- "# =====================================================================\n",
119
- "# Restart kernel, then run from this cell\n",
120
- "\n",
121
- "# =====================================================================\n",
122
- "# CELL 3: Verify NumPy Version\n",
123
- "# =====================================================================\n",
124
- "import numpy as np\n",
125
- "print(f\"✓ np: {np.__version__} - {np.__file__}\")\n",
126
- "!pip show numpy | grep Version\n",
127
- "\n",
128
- "# =====================================================================\n",
129
- "# CELL 4: Verify Roma Installation\n",
130
- "# =====================================================================\n",
131
- "try:\n",
132
- " import roma\n",
133
- " print(\"✓ roma is installed\")\n",
134
- "except ModuleNotFoundError:\n",
135
- " print(\"⚠️ roma not found, installing...\")\n",
136
- " !pip install roma\n",
137
- " import roma\n",
138
- " print(\"✓ roma installed\")"
139
- ],
140
- "metadata": {
141
- "trusted": true,
142
- "id": "XgxGC30cjLmF"
143
- },
144
- "outputs": [],
145
- "execution_count": null
146
- },
147
- {
148
- "cell_type": "code",
149
- "source": [
150
- "# =====================================================================\n",
151
- "# CELL 5: Clone Repositories\n",
152
- "# =====================================================================\n",
153
- "import os\n",
154
- "import sys\n",
155
- "\n",
156
- "# MASt3Rをクローン\n",
157
- "if not os.path.exists('/content/mast3r'):\n",
158
- " print(\"Cloning MASt3R repository...\")\n",
159
- " !git clone --recursive https://github.com/naver/mast3r.git /content/mast3r\n",
160
- " print(\"✓ MASt3R cloned\")\n",
161
- "else:\n",
162
- " print(\"✓ MASt3R already exists\")\n",
163
- "\n",
164
- "# DUSt3Rをクローン（MASt3R内に必要）\n",
165
- "if not os.path.exists('/content/mast3r/dust3r'):\n",
166
- " print(\"Cloning DUSt3R repository...\")\n",
167
- " !git clone --recursive https://github.com/naver/dust3r.git /content/mast3r/dust3r\n",
168
- " print(\"✓ DUSt3R cloned\")\n",
169
- "else:\n",
170
- " print(\"✓ DUSt3R already exists\")\n",
171
- "\n",
172
- "# ASMKをクローン\n",
173
- "if not os.path.exists('/content/asmk'):\n",
174
- " print(\"Cloning ASMK repository...\")\n",
175
- " !git clone https://github.com/jenicek/asmk.git /content/asmk\n",
176
- " print(\"✓ ASMK cloned\")\n",
177
- "else:\n",
178
- " print(\"✓ ASMK already exists\")\n",
179
- "\n",
180
- "# パスを追加\n",
181
- "sys.path.insert(0, '/content/mast3r')\n",
182
- "sys.path.insert(0, '/content/mast3r/dust3r')\n",
183
- "sys.path.insert(0, '/content/asmk')\n",
184
- "\n",
185
- "# 確認\n",
186
- "try:\n",
187
- " from dust3r.model import AsymmetricCroCo3DStereo\n",
188
- " print(\"✓ dust3r.model imported successfully\")\n",
189
- "except ImportError as e:\n",
190
- " print(f\"✗ Import error: {e}\")\n",
191
- "\n",
192
- "# croco（MASt3Rの依存関係）もクローン\n",
193
- "if not os.path.exists('/content/mast3r/croco'):\n",
194
- " print(\"Cloning CroCo repository...\")\n",
195
- " !git clone --recursive https://github.com/naver/croco.git /content/mast3r/croco\n",
196
- " print(\"✓ CroCo cloned\")\n",
197
- "\n",
198
- "# CroCo v2の依存関係\n",
199
- "if not os.path.exists('/content/mast3r/croco/models/curope'):\n",
200
- " print(\"Cloning CuRoPe...\")\n",
201
- " !git clone --recursive https://github.com/naver/curope.git /content/mast3r/croco/models/curope\n",
202
- " print(\"✓ CuRoPe cloned\")\n",
203
- "\n",
204
- "# =====================================================================\n",
205
- "# CELL 6: Clone and Build Gaussian Splatting\n",
206
- "# =====================================================================\n",
207
- "print(\"\\n\" + \"=\"*70)\n",
208
- "print(\"STEP 2: Clone Gaussian Splatting\")\n",
209
- "print(\"=\"*70)\n",
210
- "WORK_DIR = \"/content/gaussian-splatting\"\n",
211
- "\n",
212
- "import subprocess\n",
213
- "if not os.path.exists(WORK_DIR):\n",
214
- " subprocess.run([\n",
215
- " \"git\", \"clone\", \"--recursive\",\n",
216
- " \"https://github.com/graphdeco-inria/gaussian-splatting.git\",\n",
217
- " WORK_DIR\n",
218
- " ], capture_output=True)\n",
219
- " print(\"✓ Cloned\")\n",
220
- "else:\n",
221
- " print(\"✓ Already exists\")\n",
222
- "\n",
223
- "# インストールが必要なディレクトリ\n",
224
- "submodules = [\n",
225
- " \"/content/gaussian-splatting/submodules/diff-gaussian-rasterization\",\n",
226
- " \"/content/gaussian-splatting/submodules/simple-knn\"\n",
227
- "]\n",
228
- "\n",
229
- "for path in submodules:\n",
230
- " print(f\"Installing {path}...\")\n",
231
- " subprocess.run([\"pip\", \"install\", path], check=True)\n",
232
- "\n",
233
- "print(\"✓ Custom CUDA modules installed.\")\n",
234
- "\n",
235
- "# =====================================================================\n",
236
- "# CELL 7: Verify NumPy Again\n",
237
- "# =====================================================================\n",
238
- "import numpy as np\n",
239
- "print(f\"✓ np: {np.__version__} - {np.__file__}\")\n",
240
- "!pip show numpy | grep Version"
241
- ],
242
- "metadata": {
243
- "trusted": true,
244
- "id": "EF_Z8VDLjLmF"
245
- },
246
- "outputs": [],
247
- "execution_count": null
248
- },
249
- {
250
- "cell_type": "code",
251
- "source": [
252
- "# =====================================================================\n",
253
- "# CELL 8: Import Core Libraries and Configure Memory\n",
254
- "# =====================================================================\n",
255
- "import os\n",
256
- "import sys\n",
257
- "import gc\n",
258
- "import torch\n",
259
- "import numpy as np\n",
260
- "from pathlib import Path\n",
261
- "from tqdm import tqdm\n",
262
- "import torch.nn.functional as F\n",
263
- "import shutil\n",
264
- "from PIL import Image\n",
265
- "\n",
266
- "# MEMORY MANAGEMENT\n",
267
- "os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'\n",
268
- "\n",
269
- "def clear_memory():\n",
270
- " \"\"\"メモリクリア関数\"\"\"\n",
271
- " gc.collect()\n",
272
- " if torch.cuda.is_available():\n",
273
- " torch.cuda.empty_cache()\n",
274
- " torch.cuda.synchronize()\n",
275
- "\n",
276
- "# CONFIGURATION\n",
277
- "class Config:\n",
278
- " DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n",
279
- " MAST3R_WEIGHTS = \"naver/MASt3R_ViTLarge_BaseDecoder_512_catmlpdpt_metric\"\n",
280
- " DUST3R_WEIGHTS = \"naver/DUSt3R_ViTLarge_BaseDecoder_512_dpt\"\n",
281
- " RETRIEVAL_TOPK = 10\n",
282
- " IMAGE_SIZE = 224\n",
283
- "\n",
284
- "# =====================================================================\n",
285
- "# CELL 9: Image Preprocessing Functions\n",
286
- "# =====================================================================\n",
287
- "def normalize_image_sizes_biplet(input_dir, output_dir=None, size=1024):\n",
288
- " \"\"\"\n",
289
- " Generates two square crops (Left & Right or Top & Bottom)\n",
290
- " from each image in a directory.\n",
291
- " \"\"\"\n",
292
- " if output_dir is None:\n",
293
- " output_dir = input_dir + \"_biplet\"\n",
294
- "\n",
295
- " os.makedirs(output_dir, exist_ok=True)\n",
296
- "\n",
297
- " print(f\"\\n=== Generating Biplet Crops ({size}x{size}) ===\")\n",
298
- "\n",
299
- " converted_count = 0\n",
300
- " size_stats = {}\n",
301
- "\n",
302
- " for img_file in tqdm(sorted(os.listdir(input_dir)), desc=\"Creating biplets\"):\n",
303
- " if not img_file.lower().endswith(('.jpg', '.jpeg', '.png')):\n",
304
- " continue\n",
305
- "\n",
306
- " input_path = os.path.join(input_dir, img_file)\n",
307
- "\n",
308
- " try:\n",
309
- " img = Image.open(input_path)\n",
310
- " original_size = img.size\n",
311
- "\n",
312
- " size_key = f\"{original_size[0]}x{original_size[1]}\"\n",
313
- " size_stats[size_key] = size_stats.get(size_key, 0) + 1\n",
314
- "\n",
315
- " # Generate 2 crops\n",
316
- " crops = generate_two_crops(img, size)\n",
317
- "\n",
318
- " base_name, ext = os.path.splitext(img_file)\n",
319
- " for mode, cropped_img in crops.items():\n",
320
- " output_path = os.path.join(output_dir, f\"{base_name}_{mode}{ext}\")\n",
321
- " cropped_img.save(output_path, quality=95)\n",
322
- "\n",
323
- " converted_count += 1\n",
324
- "\n",
325
- " except Exception as e:\n",
326
- " print(f\" ✗ Error processing {img_file}: {e}\")\n",
327
- "\n",
328
- " print(f\"\\n✓ Biplet generation complete:\")\n",
329
- " print(f\" Source images: {converted_count}\")\n",
330
- " print(f\" Biplet crops generated: {converted_count * 2}\")\n",
331
- " print(f\" Original size distribution: {size_stats}\")\n",
332
- "\n",
333
- " return output_dir\n",
334
- "\n",
335
- "\n",
336
- "def generate_two_crops(img, size):\n",
337
- " \"\"\"\n",
338
- " Crops the image into a square and returns 2 variations\n",
339
- " \"\"\"\n",
340
- " width, height = img.size\n",
341
- " crop_size = min(width, height)\n",
342
- " crops = {}\n",
343
- "\n",
344
- " if width > height:\n",
345
- " # Landscape → Left & Right\n",
346
- " positions = {\n",
347
- " 'left': 0,\n",
348
- " 'right': width - crop_size\n",
349
- " }\n",
350
- " for mode, x_offset in positions.items():\n",
351
- " box = (x_offset, 0, x_offset + crop_size, crop_size)\n",
352
- " crops[mode] = img.crop(box).resize(\n",
353
- " (size, size),\n",
354
- " Image.Resampling.LANCZOS\n",
355
- " )\n",
356
- " else:\n",
357
- " # Portrait or Square → Top & Bottom\n",
358
- " positions = {\n",
359
- " 'top': 0,\n",
360
- " 'bottom': height - crop_size\n",
361
- " }\n",
362
- " for mode, y_offset in positions.items():\n",
363
- " box = (0, y_offset, crop_size, y_offset + crop_size)\n",
364
- " crops[mode] = img.crop(box).resize(\n",
365
- " (size, size),\n",
366
- " Image.Resampling.LANCZOS\n",
367
- " )\n",
368
- "\n",
369
- " return crops\n",
370
- "\n",
371
- "# =====================================================================\n",
372
- "# CELL 10: Image Loading Function\n",
373
- "# =====================================================================\n",
374
- "def load_images_from_directory(image_dir, max_images=200):\n",
375
- " \"\"\"ディレクトリから画像をロード\"\"\"\n",
376
- " print(f\"\\nLoading images from: {image_dir}\")\n",
377
- "\n",
378
- " valid_extensions = {'.jpg', '.jpeg', '.png', '.bmp'}\n",
379
- " image_paths = []\n",
380
- "\n",
381
- " for ext in valid_extensions:\n",
382
- " image_paths.extend(sorted(Path(image_dir).glob(f'*{ext}')))\n",
383
- " image_paths.extend(sorted(Path(image_dir).glob(f'*{ext.upper()}')))\n",
384
- "\n",
385
- " image_paths = sorted(set(str(p) for p in image_paths))\n",
386
- "\n",
387
- " if len(image_paths) > max_images:\n",
388
- " print(f\"⚠️ Limiting from {len(image_paths)} to {max_images} images\")\n",
389
- " image_paths = image_paths[:max_images]\n",
390
- "\n",
391
- " print(f\"✓ Found {len(image_paths)} images\")\n",
392
- " return image_paths"
393
- ],
394
- "metadata": {
395
- "trusted": true,
396
- "id": "_rFAsFGDjLmF"
397
- },
398
- "outputs": [],
399
- "execution_count": null
400
- },
401
- {
402
- "cell_type": "code",
403
- "source": [
404
- "# =====================================================================\n",
405
- "# CELL 11: MASt3R Model Loading\n",
406
- "# =====================================================================\n",
407
- "def load_mast3r_model(device):\n",
408
- " \"\"\"MASt3Rモデルをロード\"\"\"\n",
409
- " print(\"\\n=== Loading MASt3R Model ===\")\n",
410
- "\n",
411
- " if '/content/mast3r' not in sys.path:\n",
412
- " sys.path.insert(0, '/content/mast3r')\n",
413
- " if '/content/mast3r/dust3r' not in sys.path:\n",
414
- " sys.path.insert(0, '/content/mast3r/dust3r')\n",
415
- "\n",
416
- " from dust3r.model import AsymmetricCroCo3DStereo\n",
417
- "\n",
418
- " try:\n",
419
- " print(f\"Attempting to load: {Config.MAST3R_WEIGHTS}\")\n",
420
- " model = AsymmetricCroCo3DStereo.from_pretrained(Config.MAST3R_WEIGHTS).to(device)\n",
421
- " print(\"✓ Loaded MASt3R model\")\n",
422
- " except Exception as e:\n",
423
- " print(f\"⚠️ Failed to load MASt3R: {e}\")\n",
424
- " print(f\"Trying DUSt3R instead: {Config.DUST3R_WEIGHTS}\")\n",
425
- " model = AsymmetricCroCo3DStereo.from_pretrained(Config.DUST3R_WEIGHTS).to(device)\n",
426
- " print(\"✓ Loaded DUSt3R model as fallback\")\n",
427
- "\n",
428
- " model.eval()\n",
429
- " print(f\"✓ Model loaded on {device}\")\n",
430
- " return model\n",
431
- "\n",
432
- "# =====================================================================\n",
433
- "# CELL 12: Feature Extraction (FIXED)\n",
434
- "# =====================================================================\n",
435
- "def extract_mast3r_features(model, image_paths, device, batch_size=1):\n",
436
- " \"\"\"MASt3Rモデルを使用して特徴量を抽出（修正版）\"\"\"\n",
437
- " print(\"\\n=== Extracting MASt3R Features ===\")\n",
438
- " from dust3r.utils.image import load_images\n",
439
- " from dust3r.inference import inference\n",
440
- "\n",
441
- " all_features = []\n",
442
- "\n",
443
- " for i in tqdm(range(len(image_paths)), desc=\"Features\"):\n",
444
- " img_path = image_paths[i]\n",
445
- "\n",
446
- " # 同じ画像を2回ロード（ペアとして）\n",
447
- " images = load_images([img_path, img_path], size=Config.IMAGE_SIZE)\n",
448
- " pairs = [(images[0], images[1])]\n",
449
- "\n",
450
- " with torch.no_grad():\n",
451
- " output = inference(pairs, model, device, batch_size=1)\n",
452
- "\n",
453
- " try:\n",
454
- " # outputから特徴量を抽出（修正版）\n",
455
- " if isinstance(output, dict):\n",
456
- " if 'pred1' in output:\n",
457
- " pred1 = output['pred1']\n",
458
- " if isinstance(pred1, dict):\n",
459
- " # 'desc'または'conf'を優先的に使用\n",
460
- " if 'desc' in pred1:\n",
461
- " desc = pred1['desc']\n",
462
- " elif 'conf' in pred1:\n",
463
- " desc = pred1['conf']\n",
464
- " elif 'pts3d' in pred1:\n",
465
- " desc = pred1['pts3d']\n",
466
- " else:\n",
467
- " desc = list(pred1.values())[0]\n",
468
- " else:\n",
469
- " desc = pred1\n",
470
- " elif 'view1' in output:\n",
471
- " view1 = output['view1']\n",
472
- " if isinstance(view1, dict):\n",
473
- " desc = view1.get('desc', view1.get('conf', view1.get('pts3d', list(view1.values())[0])))\n",
474
- " else:\n",
475
- " desc = view1\n",
476
- " else:\n",
477
- " desc = list(output.values())[0]\n",
478
- " elif isinstance(output, tuple) and len(output) == 2:\n",
479
- " view1, view2 = output\n",
480
- " if isinstance(view1, dict):\n",
481
- " desc = view1.get('desc', view1.get('conf', view1.get('pts3d', list(view1.values())[0])))\n",
482
- " else:\n",
483
- " desc = view1\n",
484
- " elif isinstance(output, list):\n",
485
- " item = output[0]\n",
486
- " if isinstance(item, dict):\n",
487
- " desc = item.get('desc', item.get('conf', item.get('pts3d', list(item.values())[0])))\n",
488
- " else:\n",
489
- " desc = item\n",
490
- " else:\n",
491
- " desc = output\n",
492
- "\n",
493
- " # テンソルをCPUに移動して保存\n",
494
- " if isinstance(desc, torch.Tensor):\n",
495
- " desc = desc.detach().cpu()\n",
496
- "\n",
497
- " # 4次元の場合はbatch次元を削除\n",
498
- " if desc.dim() == 4:\n",
499
- " desc = desc.squeeze(0)\n",
500
- "\n",
501
- " # 特徴量の次元が小さすぎる場合（RGB画像など）は平均プーリング\n",
502
- " if desc.shape[-1] < 16:\n",
503
- " # [H, W, 3] -> [H, W, 64] に拡張\n",
504
- " desc = desc.unsqueeze(-1).repeat(1, 1, 1, 64 // desc.shape[-1]).reshape(desc.shape[0], desc.shape[1], -1)\n",
505
- "\n",
506
- " all_features.append(desc)\n",
507
- "\n",
508
- " except Exception as e:\n",
509
- " print(f\"⚠️ Error extracting features for image {i}: {e}\")\n",
510
- " # デフォルト特徴量\n",
511
- " all_features.append(torch.zeros((Config.IMAGE_SIZE, Config.IMAGE_SIZE, 64)))\n",
512
- "\n",
513
- " # メモリクリア\n",
514
- " del output, images, pairs\n",
515
- " if i % 10 == 0:\n",
516
- " torch.cuda.empty_cache()\n",
517
- "\n",
518
- " print(f\"✓ Extracted features for {len(all_features)} images\")\n",
519
- " if all_features:\n",
520
- " first_feat = all_features[0]\n",
521
- " if isinstance(first_feat, torch.Tensor):\n",
522
- " print(f\" Feature shape: {first_feat.shape}\")\n",
523
- "\n",
524
- " return all_features\n",
525
- "\n",
526
- "# =====================================================================\n",
527
- "# CELL 13: ASMK Similarity Computation (FIXED)\n",
528
- "# =====================================================================\n",
529
- "def compute_asmk_similarity(features, codebook=None):\n",
530
- " \"\"\"ASMKを使用して類似度行列を計算（修正版）\"\"\"\n",
531
- " print(\"\\n=== Computing ASMK Similarity ===\")\n",
532
- "\n",
533
- " n_images = len(features)\n",
534
- " similarity_matrix = np.zeros((n_images, n_images), dtype=np.float32)\n",
535
- "\n",
536
- " # 各特徴量をグローバル記述子に変換\n",
537
- " global_features = []\n",
538
- "\n",
539
- " for feat in features:\n",
540
- " if isinstance(feat, dict):\n",
541
- " for key in ['desc', 'conf', 'pts3d']:\n",
542
- " if key in feat:\n",
543
- " feat = feat[key]\n",
544
- " break\n",
545
- "\n",
546
- " if isinstance(feat, torch.Tensor):\n",
547
- " feat = feat.cpu().numpy()\n",
548
- "\n",
549
- " if isinstance(feat, np.ndarray):\n",
550
- " if feat.ndim == 3: # [H, W, C]\n",
551
- " feat_flat = feat.reshape(-1, feat.shape[-1])\n",
552
- " elif feat.ndim == 2: # [N, C]\n",
553
- " feat_flat = feat\n",
554
- " else:\n",
555
- " feat_flat = feat.reshape(-1, max(feat.shape))\n",
556
- "\n",
557
- " global_desc = np.mean(feat_flat, axis=0)\n",
558
- " global_features.append(global_desc)\n",
559
- " else:\n",
560
- " # ダミー特徴量\n",
561
- " global_features.append(np.zeros(64))\n",
562
- "\n",
563
- " global_features = np.stack(global_features)\n",
564
- " feature_dim = global_features.shape[1]\n",
565
- "\n",
566
- " print(f\"Global features shape: {global_features.shape}\")\n",
567
- "\n",
568
- " # コサイン類似度を使用\n",
569
- " global_features_norm = global_features / (np.linalg.norm(global_features, axis=1, keepdims=True) + 1e-8)\n",
570
- " similarity_matrix = global_features_norm @ global_features_norm.T\n",
571
- "\n",
572
- " np.fill_diagonal(similarity_matrix, -1)\n",
573
- "\n",
574
- " print(f\"Similarity matrix shape: {similarity_matrix.shape}\")\n",
575
- " print(f\"Similarity range: [{similarity_matrix.min():.3f}, {similarity_matrix.max():.3f}]\")\n",
576
- "\n",
577
- " return similarity_matrix\n",
578
- "\n",
579
- "\n",
580
- "def build_pairs_from_similarity(similarity_matrix, top_k=10):\n",
581
- " \"\"\"類似度行列からペアを構築\"\"\"\n",
582
- " n_images = similarity_matrix.shape[0]\n",
583
- " pairs = []\n",
584
- "\n",
585
- " for i in range(n_images):\n",
586
- " similarities = similarity_matrix[i]\n",
587
- " top_indices = np.argsort(similarities)[::-1][:top_k]\n",
588
- "\n",
589
- " for j in top_indices:\n",
590
- " if j > i:\n",
591
- " pairs.append((i, j))\n",
592
- "\n",
593
- " pairs = list(set(pairs))\n",
594
- " print(f\"✓ Built {len(pairs)} unique pairs\")\n",
595
- "\n",
596
- " return pairs\n",
597
- "\n",
598
- "\n",
599
- "def get_image_pairs_asmk(image_paths, max_pairs=100):\n",
600
- " \"\"\"ASMKを使用して画像ペアを取得\"\"\"\n",
601
- " print(\"\\n=== Getting Image Pairs with ASMK ===\")\n",
602
- "\n",
603
- " device = Config.DEVICE\n",
604
- " model = load_mast3r_model(device)\n",
605
- " features = extract_mast3r_features(model, image_paths, device)\n",
606
- " similarity_matrix = compute_asmk_similarity(features)\n",
607
- " pairs = build_pairs_from_similarity(similarity_matrix, Config.RETRIEVAL_TOPK)\n",
608
- "\n",
609
- " # モデルを解放\n",
610
- " del model\n",
611
- " clear_memory()\n",
612
- "\n",
613
- " if len(pairs) > max_pairs:\n",
614
- " pairs = pairs[:max_pairs]\n",
615
- " print(f\"Limited to {max_pairs} pairs\")\n",
616
- "\n",
617
- " return pairs"
618
- ],
619
- "metadata": {
620
- "trusted": true,
621
- "id": "qo0mGj_5jLmG"
622
- },
623
- "outputs": [],
624
- "execution_count": null
625
- },
626
- {
627
- "cell_type": "code",
628
- "source": [
629
- "# =====================================================================\n",
630
- "# CELL 14: MASt3R Reconstruction\n",
631
- "# =====================================================================\n",
632
- "def run_mast3r_pairs(model, image_paths, pairs, device, batch_size=1):\n",
633
- " \"\"\"MASt3Rでペア画像を処理（メモリ最適化版）\"\"\"\n",
634
- " print(\"\\n=== Running MASt3R Reconstruction ===\")\n",
635
- " from dust3r.inference import inference\n",
636
- " from dust3r.cloud_opt import global_aligner, GlobalAlignerMode\n",
637
- " from dust3r.utils.image import load_images\n",
638
- "\n",
639
- " # ペアを制限\n",
640
- " max_pairs_for_memory = 50\n",
641
- " if len(pairs) > max_pairs_for_memory:\n",
642
- " print(f\"⚠️ Limiting pairs from {len(pairs)} to {max_pairs_for_memory} for memory\")\n",
643
- " pairs = pairs[:max_pairs_for_memory]\n",
644
- "\n",
645
- " # ペアから画像インデックスを取得\n",
646
- " pair_indices = []\n",
647
- " for i, j in pairs:\n",
648
- " pair_indices.extend([i, j])\n",
649
- " unique_indices = sorted(set(pair_indices))\n",
650
- "\n",
651
- " selected_paths = [image_paths[i] for i in unique_indices]\n",
652
- " print(f\"Selected {len(selected_paths)} unique images from {len(pairs)} pairs\")\n",
653
- "\n",
654
- " # 画像をロード\n",
655
- " images = load_images(selected_paths, size=Config.IMAGE_SIZE)\n",
656
- " clear_memory()\n",
657
- "\n",
658
- " # インデックスマッピング\n",
659
- " index_map = {old_idx: new_idx for new_idx, old_idx in enumerate(unique_indices)}\n",
660
- "\n",
661
- " # ペア画像リストを作成\n",
662
- " image_pairs = []\n",
663
- " for i, j in pairs:\n",
664
- " new_i = index_map[i]\n",
665
- " new_j = index_map[j]\n",
666
- " image_pairs.append((images[new_i], images[new_j]))\n",
667
- "\n",
668
- " print(f\"Created {len(image_pairs)} image pairs\")\n",
669
- " clear_memory()\n",
670
- "\n",
671
- " # 推論を実行\n",
672
- " print(f\"Running inference on {len(image_pairs)} pairs...\")\n",
673
- " with torch.no_grad():\n",
674
- " output = inference(image_pairs, model, device, batch_size=batch_size)\n",
675
- "\n",
676
- " print(f\"✓ Processed {len(output)} predictions\")\n",
677
- " clear_memory()\n",
678
- "\n",
679
- " # Global alignment\n",
680
- " scene = global_aligner(\n",
681
- " dust3r_output=output,\n",
682
- " device=device,\n",
683
- " mode=GlobalAlignerMode.PointCloudOptimizer,\n",
684
- " verbose=True\n",
685
- " )\n",
686
- "\n",
687
- " clear_memory()\n",
688
- "\n",
689
- " print(\"Running global alignment...\")\n",
690
- " try:\n",
691
- " loss = scene.compute_global_alignment(\n",
692
- " init=\"mst\",\n",
693
- " niter=50,\n",
694
- " schedule='cosine',\n",
695
- " lr=0.01\n",
696
- " )\n",
697
- " print(f\"✓ Alignment complete (loss: {loss:.6f})\")\n",
698
- " except RuntimeError as e:\n",
699
- " if \"out of memory\" in str(e).lower():\n",
700
- " print(\"⚠️ OOM during alignment, trying with fewer iterations...\")\n",
701
- " clear_memory()\n",
702
- " loss = scene.compute_global_alignment(\n",
703
- " init=\"mst\",\n",
704
- " niter=20,\n",
705
- " schedule='cosine',\n",
706
- " lr=0.01\n",
707
- " )\n",
708
- " print(f\"✓ Alignment complete with reduced iterations (loss: {loss:.6f})\")\n",
709
- " else:\n",
710
- " raise\n",
711
- "\n",
712
- " clear_memory()\n",
713
- " return scene, images\n",
714
- "\n",
715
- "# =====================================================================\n",
716
- "# CELL 15: Camera Parameter Extraction\n",
717
- "# =====================================================================\n",
718
- "def extract_camera_params_process2(scene, image_paths, conf_threshold=1.5):\n",
719
- " \"\"\"sceneからカメラパラメータと3D点を抽出\"\"\"\n",
720
- " print(\"\\n=== Extracting Camera Parameters ===\")\n",
721
- "\n",
722
- " cameras_dict = {}\n",
723
- " all_pts3d = []\n",
724
- " all_confidence = []\n",
725
- "\n",
726
- " try:\n",
727
- " if hasattr(scene, 'get_im_poses'):\n",
728
- " poses = scene.get_im_poses()\n",
729
- " elif hasattr(scene, 'im_poses'):\n",
730
- " poses = scene.im_poses\n",
731
- " else:\n",
732
- " poses = None\n",
733
- "\n",
734
- " if hasattr(scene, 'get_focals'):\n",
735
- " focals = scene.get_focals()\n",
736
- " elif hasattr(scene, 'im_focals'):\n",
737
- " focals = scene.im_focals\n",
738
- " else:\n",
739
- " focals = None\n",
740
- "\n",
741
- " if hasattr(scene, 'get_principal_points'):\n",
742
- " pps = scene.get_principal_points()\n",
743
- " elif hasattr(scene, 'im_pp'):\n",
744
- " pps = scene.im_pp\n",
745
- " else:\n",
746
- " pps = None\n",
747
- " except Exception as e:\n",
748
- " print(f\"⚠️ Error getting camera parameters: {e}\")\n",
749
- " poses = None\n",
750
- " focals = None\n",
751
- " pps = None\n",
752
- "\n",
753
- " n_images = min(len(poses) if poses is not None else len(image_paths), len(image_paths))\n",
754
- "\n",
755
- " for idx in range(n_images):\n",
756
- " img_name = os.path.basename(image_paths[idx])\n",
757
- "\n",
758
- " try:\n",
759
- " # Poseを取得\n",
760
- " if poses is not None and idx < len(poses):\n",
761
- " pose = poses[idx]\n",
762
- " if isinstance(pose, torch.Tensor):\n",
763
- " pose = pose.detach().cpu().numpy()\n",
764
- " if not isinstance(pose, np.ndarray) or pose.shape != (4, 4):\n",
765
- " pose = np.eye(4)\n",
766
- " else:\n",
767
- " pose = np.eye(4)\n",
768
- "\n",
769
- " # Focalを取得\n",
770
- " if focals is not None and idx < len(focals):\n",
771
- " focal = focals[idx]\n",
772
- " if isinstance(focal, torch.Tensor):\n",
773
- " focal = focal.detach().cpu().item()\n",
774
- " else:\n",
775
- " focal = float(focal)\n",
776
- " else:\n",
777
- " focal = 1000.0\n",
778
- "\n",
779
- " # Principal pointを取得\n",
780
- " if pps is not None and idx < len(pps):\n",
781
- " pp = pps[idx]\n",
782
- " if isinstance(pp, torch.Tensor):\n",
783
- " pp = pp.detach().cpu().numpy()\n",
784
- " else:\n",
785
- " pp = np.array([112.0, 112.0])\n",
786
- "\n",
787
- " # カメラパラメータを保存\n",
788
- " cameras_dict[img_name] = {\n",
789
- " 'focal': focal,\n",
790
- " 'pp': pp,\n",
791
- " 'pose': pose,\n",
792
- " 'width': Config.IMAGE_SIZE * 4,\n",
793
- " 'height': Config.IMAGE_SIZE * 4\n",
794
- " }\n",
795
- "\n",
796
- " # 3D点を取得\n",
797
- " if hasattr(scene, 'im_pts3d') and idx < len(scene.im_pts3d):\n",
798
- " pts3d_img = scene.im_pts3d[idx]\n",
799
- " elif hasattr(scene, 'get_pts3d'):\n",
800
- " pts3d_all = scene.get_pts3d()\n",
801
- " if idx < len(pts3d_all):\n",
802
- " pts3d_img = pts3d_all[idx]\n",
803
- " else:\n",
804
- " pts3d_img = None\n",
805
- " else:\n",
806
- " pts3d_img = None\n",
807
- "\n",
808
- " # Confidenceを取得\n",
809
- " if hasattr(scene, 'im_conf') and idx < len(scene.im_conf):\n",
810
- " conf_img = scene.im_conf[idx]\n",
811
- " elif hasattr(scene, 'get_conf'):\n",
812
- " conf_all = scene.get_conf()\n",
813
- " if idx < len(conf_all):\n",
814
- " conf_img = conf_all[idx]\n",
815
- " else:\n",
816
- " conf_img = None\n",
817
- " else:\n",
818
- " conf_img = None\n",
819
- "\n",
820
- " # 3D点とconfidenceを処理\n",
821
- " if pts3d_img is not None:\n",
822
- " if isinstance(pts3d_img, torch.Tensor):\n",
823
- " pts3d_img = pts3d_img.detach().cpu().numpy()\n",
824
- "\n",
825
- " if pts3d_img.ndim == 3:\n",
826
- " pts3d_flat = pts3d_img.reshape(-1, 3)\n",
827
- " else:\n",
828
- " pts3d_flat = pts3d_img\n",
829
- "\n",
830
- " all_pts3d.append(pts3d_flat)\n",
831
- "\n",
832
- " # confidenceを処理\n",
833
- " if conf_img is not None:\n",
834
- " if isinstance(conf_img, list):\n",
835
- " conf_img = np.array(conf_img)\n",
836
- " elif isinstance(conf_img, torch.Tensor):\n",
837
- " conf_img = conf_img.detach().cpu().numpy()\n",
838
- "\n",
839
- " if conf_img.ndim > 1:\n",
840
- " conf_flat = conf_img.reshape(-1)\n",
841
- " else:\n",
842
- " conf_flat = conf_img\n",
843
- "\n",
844
- " if len(conf_flat) != len(pts3d_flat):\n",
845
- " conf_flat = np.ones(len(pts3d_flat))\n",
846
- "\n",
847
- " all_confidence.append(conf_flat)\n",
848
- " else:\n",
849
- " all_confidence.append(np.ones(len(pts3d_flat)))\n",
850
- "\n",
851
- " except Exception as e:\n",
852
- " print(f\"⚠️ Error processing image {idx} ({img_name}): {e}\")\n",
853
- " cameras_dict[img_name] = {\n",
854
- " 'focal': 1000.0,\n",
855
- " 'pp': np.array([112.0, 112.0]),\n",
856
- " 'pose': np.eye(4),\n",
857
- " 'width': Config.IMAGE_SIZE * 4,\n",
858
- " 'height': Config.IMAGE_SIZE * 4\n",
859
- " }\n",
860
- " continue\n",
861
- "\n",
862
- " # 全3D点を結合\n",
863
- " if all_pts3d:\n",
864
- " pts3d = np.vstack(all_pts3d)\n",
865
- " confidence = np.concatenate(all_confidence)\n",
866
- " else:\n",
867
- " pts3d = np.zeros((0, 3))\n",
868
- " confidence = np.zeros(0)\n",
869
- "\n",
870
- " print(f\"✓ Extracted camera parameters for {len(cameras_dict)} images\")\n",
871
- " print(f\"✓ Total 3D points: {len(pts3d)}\")\n",
872
- "\n",
873
- " # Confidenceでフィルタリング\n",
874
- " if len(confidence) > 0:\n",
875
- " valid_mask = confidence > conf_threshold\n",
876
- " pts3d = pts3d[valid_mask]\n",
877
- " confidence = confidence[valid_mask]\n",
878
- " print(f\"✓ After confidence filtering (>{conf_threshold}): {len(pts3d)} points\")\n",
879
- "\n",
880
- " return cameras_dict, pts3d, confidence"
881
- ],
882
- "metadata": {
883
- "trusted": true,
884
- "id": "bCXpdw83jLmG"
885
- },
886
- "outputs": [],
887
- "execution_count": null
888
- },
889
- {
890
- "cell_type": "code",
891
- "source": [
892
- "# =====================================================================\n",
893
- "# CELL 16: COLMAP Export Functions\n",
894
- "# =====================================================================\n",
895
- "import struct\n",
896
- "from scipy.spatial.transform import Rotation as R\n",
897
- "\n",
898
- "def write_colmap_sparse(cameras_dict, pts3d, confidence, image_paths, output_dir):\n",
899
- " \"\"\"COLMAP sparse形式をバイナリファイルで出力\"\"\"\n",
900
- " os.makedirs(output_dir, exist_ok=True)\n",
901
- "\n",
902
- " if not cameras_dict:\n",
903
- " raise ValueError(\"cameras_dict is empty\")\n",
904
- "\n",
905
- " first_key = list(cameras_dict.keys())[0]\n",
906
- " first_cam = cameras_dict[first_key]\n",
907
- "\n",
908
- " w = int(first_cam.get('width', 1920))\n",
909
- " h = int(first_cam.get('height', 1080))\n",
910
- " focal = float(first_cam.get('focal', max(w, h) * 1.2))\n",
911
- " cx = w / 2.0\n",
912
- " cy = h / 2.0\n",
913
- "\n",
914
- " # cameras.bin\n",
915
- " cameras_file = os.path.join(output_dir, 'cameras.bin')\n",
916
- " with open(cameras_file, 'wb') as f:\n",
917
- " f.write(struct.pack('Q', 1))\n",
918
- " camera_id = 1\n",
919
- " model_id = 1 # PINHOLE\n",
920
- " f.write(struct.pack('i', camera_id))\n",
921
- " f.write(struct.pack('i', model_id))\n",
922
- " f.write(struct.pack('Q', w))\n",
923
- " f.write(struct.pack('Q', h))\n",
924
- " f.write(struct.pack('d', focal))\n",
925
- " f.write(struct.pack('d', focal))\n",
926
- " f.write(struct.pack('d', cx))\n",
927
- " f.write(struct.pack('d', cy))\n",
928
- "\n",
929
- " print(f\"✓ Written cameras.bin\")\n",
930
- "\n",
931
- " # images.bin\n",
932
- " images_file = os.path.join(output_dir, 'images.bin')\n",
933
- " with open(images_file, 'wb') as f:\n",
934
- " f.write(struct.pack('Q', len(image_paths)))\n",
935
- "\n",
936
- " for i, img_path in enumerate(image_paths):\n",
937
- " img_name = os.path.basename(img_path)\n",
938
- "\n",
939
- " cam_info = cameras_dict.get(img_name)\n",
940
- " if cam_info is None:\n",
941
- " pose = np.eye(4)\n",
942
- " else:\n",
943
- " pose = cam_info['pose']\n",
944
- "\n",
945
- " try:\n",
946
- " w2c = np.linalg.inv(pose)\n",
947
- " except np.linalg.LinAlgError:\n",
948
- " w2c = np.eye(4)\n",
949
- "\n",
950
- " rot_mat = w2c[:3, :3]\n",
951
- " tvec = w2c[:3, 3]\n",
952
- " quat = R.from_matrix(rot_mat).as_quat()\n",
953
- " qw, qx, qy, qz = quat[3], quat[0], quat[1], quat[2]\n",
954
- "\n",
955
- " image_id = i + 1\n",
956
- " f.write(struct.pack('i', image_id))\n",
957
- " f.write(struct.pack('d', qw))\n",
958
- " f.write(struct.pack('d', qx))\n",
959
- " f.write(struct.pack('d', qy))\n",
960
- " f.write(struct.pack('d', qz))\n",
961
- " f.write(struct.pack('d', tvec[0]))\n",
962
- " f.write(struct.pack('d', tvec[1]))\n",
963
- " f.write(struct.pack('d', tvec[2]))\n",
964
- " f.write(struct.pack('i', 1))\n",
965
- " img_name_bytes = img_name.encode('utf-8') + b'\\x00'\n",
966
- " f.write(img_name_bytes)\n",
967
- " f.write(struct.pack('Q', 0))\n",
968
- "\n",
969
- " print(f\"✓ Written images.bin ({len(image_paths)} images)\")\n",
970
- "\n",
971
- " # points3D.bin\n",
972
- " points_file = os.path.join(output_dir, 'points3D.bin')\n",
973
- " with open(points_file, 'wb') as f:\n",
974
- " f.write(struct.pack('Q', len(pts3d)))\n",
975
- "\n",
976
- " for point_id, point in enumerate(pts3d, start=1):\n",
977
- " f.write(struct.pack('Q', point_id))\n",
978
- " f.write(struct.pack('d', point[0]))\n",
979
- " f.write(struct.pack('d', point[1]))\n",
980
- " f.write(struct.pack('d', point[2]))\n",
981
- " f.write(struct.pack('B', 255))\n",
982
- " f.write(struct.pack('B', 255))\n",
983
- " f.write(struct.pack('B', 255))\n",
984
- " f.write(struct.pack('d', 0.0))\n",
985
- " f.write(struct.pack('Q', 0))\n",
986
- "\n",
987
- " print(f\"✓ Written points3D.bin ({len(pts3d)} points)\")\n",
988
- "\n",
989
- " # テキスト形式も出力\n",
990
- " write_text_versions(cameras_dict, pts3d, image_paths, output_dir, w, h, focal, cx, cy)\n",
991
- "\n",
992
- " print(f\"\\n✓ COLMAP sparse reconstruction saved\")\n",
993
- " return output_dir\n",
994
- "\n",
995
- "\n",
996
- "def write_text_versions(cameras_dict, pts3d, image_paths, output_dir, w, h, focal, cx, cy):\n",
997
- " \"\"\"テキスト形式を出力\"\"\"\n",
998
- "\n",
999
- " # cameras.txt\n",
1000
- " with open(os.path.join(output_dir, 'cameras.txt'), 'w') as file:\n",
1001
- " file.write(\"# Camera list with one line of data per camera:\\n\")\n",
1002
- " file.write(\"# CAMERA_ID, MODEL, WIDTH, HEIGHT, PARAMS[]\\n\")\n",
1003
- " file.write(f\"1 PINHOLE {w} {h} {focal} {focal} {cx} {cy}\\n\")\n",
1004
- "\n",
1005
- " # images.txt\n",
1006
- " with open(os.path.join(output_dir, 'images.txt'), 'w') as file:\n",
1007
- " file.write(\"# Image list with two lines of data per image:\\n\")\n",
1008
- " file.write(\"# IMAGE_ID, QW, QX, QY, QZ, TX, TY, TZ, CAMERA_ID, NAME\\n\")\n",
1009
- " file.write(\"# POINTS2D[] as (X, Y, POINT3D_ID)\\n\")\n",
1010
- "\n",
1011
- " for i, img_path in enumerate(image_paths):\n",
1012
- " img_name = os.path.basename(img_path)\n",
1013
- " cam_info = cameras_dict.get(img_name)\n",
1014
- "\n",
1015
- " if cam_info is None:\n",
1016
- " pose = np.eye(4)\n",
1017
- " else:\n",
1018
- " pose = cam_info['pose']\n",
1019
- "\n",
1020
- " try:\n",
1021
- " w2c = np.linalg.inv(pose)\n",
1022
- " except np.linalg.LinAlgError:\n",
1023
- " w2c = np.eye(4)\n",
1024
- "\n",
1025
- " rot_mat = w2c[:3, :3]\n",
1026
- " tvec = w2c[:3, 3]\n",
1027
- " quat = R.from_matrix(rot_mat).as_quat()\n",
1028
- " qw, qx, qy, qz = quat[3], quat[0], quat[1], quat[2]\n",
1029
- "\n",
1030
- " image_id = i + 1\n",
1031
- " file.write(f\"{image_id} {qw} {qx} {qy} {qz} {tvec[0]} {tvec[1]} {tvec[2]} 1 {img_name}\\n\")\n",
1032
- " file.write(\"\\n\")\n",
1033
- "\n",
1034
- " # points3D.txt\n",
1035
- " with open(os.path.join(output_dir, 'points3D.txt'), 'w') as file:\n",
1036
- " file.write(\"# 3D point list with one line of data per point:\\n\")\n",
1037
- " file.write(\"# POINT3D_ID, X, Y, Z, R, G, B, ERROR, TRACK[]\\n\")\n",
1038
- "\n",
1039
- " for point_id, point in enumerate(pts3d, start=1):\n",
1040
- " file.write(f\"{point_id} {point[0]} {point[1]} {point[2]} 255 255 255 0.0\\n\")\n",
1041
- "\n",
1042
- "# =====================================================================\n",
1043
- "# CELL 17: Gaussian Splatting Runner\n",
1044
- "# =====================================================================\n",
1045
- "def run_gaussian_splatting(source_dir, output_dir, iterations=30000):\n",
1046
- " \"\"\"Gaussian Splattingを実行\"\"\"\n",
1047
- " print(\"\\n=== Running Gaussian Splatting ===\")\n",
1048
- "\n",
1049
- " os.makedirs(output_dir, exist_ok=True)\n",
1050
- "\n",
1051
- " cmd = [\n",
1052
- " \"python\", \"/content/gaussian-splatting/train.py\",\n",
1053
- " \"-s\", source_dir,\n",
1054
- " \"-m\", output_dir,\n",
1055
- " \"--iterations\", str(iterations),\n",
1056
- " \"--eval\"\n",
1057
- " ]\n",
1058
- "\n",
1059
- " print(f\"Command: {' '.join(cmd)}\")\n",
1060
- " print(f\" Source: {source_dir}\")\n",
1061
- " print(f\" Output: {output_dir}\")\n",
1062
- "\n",
1063
- " result = subprocess.run(cmd, capture_output=False, text=True)\n",
1064
- "\n",
1065
- " if result.returncode == 0:\n",
1066
- " print(f\"\\n✓ Gaussian Splatting complete\")\n",
1067
- "\n",
1068
- " point_cloud_dir = os.path.join(output_dir, \"point_cloud\")\n",
1069
- " if os.path.exists(point_cloud_dir):\n",
1070
- " print(f\"\\n✓ Point cloud directory found: {point_cloud_dir}\")\n",
1071
- "\n",
1072
- " for item in sorted(os.listdir(point_cloud_dir)):\n",
1073
- " item_path = os.path.join(point_cloud_dir, item)\n",
1074
- " if os.path.isdir(item_path) and item.startswith(\"iteration_\"):\n",
1075
- " ply_file = os.path.join(item_path, \"point_cloud.ply\")\n",
1076
- " if os.path.exists(ply_file):\n",
1077
- " file_size = os.path.getsize(ply_file) / (1024 * 1024)\n",
1078
- " print(f\" ✓ {item}/point_cloud.ply ({file_size:.2f} MB)\")\n",
1079
- " else:\n",
1080
- " print(f\"\\n✗ Gaussian Splatting failed with return code {result.returncode}\")\n",
1081
- "\n",
1082
- " return output_dir"
1083
- ],
1084
- "metadata": {
1085
- "trusted": true,
1086
- "id": "1yyRoxHKjLmH"
1087
- },
1088
- "outputs": [],
1089
- "execution_count": null
1090
- },
1091
- {
1092
- "cell_type": "code",
1093
- "source": [
1094
- "# =====================================================================\n",
1095
- "# CELL 18: Main Pipeline\n",
1096
- "# =====================================================================\n",
1097
- "def main_pipeline(image_dir, output_dir, square_size=1024, iterations=30000,\n",
1098
- " max_images=200, max_pairs=100, max_points=500000,\n",
1099
- " conf_threshold=1.5, preprocess_mode='none'):\n",
1100
- " \"\"\"メインパイプライン（修正版）\"\"\"\n",
1101
- "\n",
1102
- " # STEP 0: Image Preprocessing\n",
1103
- " if preprocess_mode == 'biplet':\n",
1104
- " print(\"=\"*70)\n",
1105
- " print(\"STEP 0: Image Preprocessing (Biplet Crops)\")\n",
1106
- " print(\"=\"*70)\n",
1107
- "\n",
1108
- " temp_biplet_dir = os.path.join(output_dir, \"temp_biplet\")\n",
1109
- " biplet_dir = normalize_image_sizes_biplet(image_dir, temp_biplet_dir, size=square_size)\n",
1110
- "\n",
1111
- " images_dir = os.path.join(output_dir, \"images\")\n",
1112
- " os.makedirs(images_dir, exist_ok=True)\n",
1113
- "\n",
1114
- " biplet_suffixes = ['_left', '_right', '_top', '_bottom']\n",
1115
- " copied_count = 0\n",
1116
- "\n",
1117
- " for img_file in os.listdir(temp_biplet_dir):\n",
1118
- " if any(suffix in img_file for suffix in biplet_suffixes):\n",
1119
- " src = os.path.join(temp_biplet_dir, img_file)\n",
1120
- " dst = os.path.join(images_dir, img_file)\n",
1121
- " shutil.copy2(src, dst)\n",
1122
- " copied_count += 1\n",
1123
- "\n",
1124
- " print(f\"✓ Copied {copied_count} biplet images to {images_dir}\")\n",
1125
- "\n",
1126
- " original_images_dir = os.path.join(output_dir, \"original_images\")\n",
1127
- " os.makedirs(original_images_dir, exist_ok=True)\n",
1128
- "\n",
1129
- " original_count = 0\n",
1130
- " valid_extensions = ('.jpg', '.jpeg', '.png', '.bmp')\n",
1131
- " for img_file in os.listdir(image_dir):\n",
1132
- " if img_file.lower().endswith(valid_extensions):\n",
1133
- " src = os.path.join(image_dir, img_file)\n",
1134
- " dst = os.path.join(original_images_dir, img_file)\n",
1135
- " shutil.copy2(src, dst)\n",
1136
- " original_count += 1\n",
1137
- "\n",
1138
- " print(f\"✓ Saved {original_count} original images to {original_images_dir}\")\n",
1139
- " shutil.rmtree(temp_biplet_dir)\n",
1140
- " image_dir = images_dir\n",
1141
- " clear_memory()\n",
1142
- " else:\n",
1143
- " images_dir = os.path.join(output_dir, \"images\")\n",
1144
- " if not os.path.exists(images_dir):\n",
1145
- " print(\"=\"*70)\n",
1146
- " print(\"STEP 0: Copying images to output directory\")\n",
1147
- " print(\"=\"*70)\n",
1148
- " shutil.copytree(image_dir, images_dir)\n",
1149
- " print(f\"✓ Copied images to {images_dir}\")\n",
1150
- " image_dir = images_dir\n",
1151
- "\n",
1152
- " # STEP 1: Loading Images\n",
1153
- " print(\"\\n\" + \"=\"*70)\n",
1154
- " print(\"STEP 1: Loading and Preparing Images\")\n",
1155
- " print(\"=\"*70)\n",
1156
- "\n",
1157
- " image_paths = load_images_from_directory(image_dir, max_images=max_images)\n",
1158
- " print(f\"Loaded {len(image_paths)} images\")\n",
1159
- " clear_memory()\n",
1160
- "\n",
1161
- " # STEP 2: Image Pair Selection\n",
1162
- " print(\"\\n\" + \"=\"*70)\n",
1163
- " print(\"STEP 2: Image Pair Selection\")\n",
1164
- " print(\"=\"*70)\n",
1165
- "\n",
1166
- " max_pairs = min(max_pairs, 50)\n",
1167
- " pairs = get_image_pairs_asmk(image_paths, max_pairs=max_pairs)\n",
1168
- " print(f\"Selected {len(pairs)} image pairs\")\n",
1169
- " clear_memory()\n",
1170
- "\n",
1171
- " # STEP 3: MASt3R 3D Reconstruction\n",
1172
- " print(\"\\n\" + \"=\"*70)\n",
1173
- " print(\"STEP 3: MASt3R 3D Reconstruction\")\n",
1174
- " print(\"=\"*70)\n",
1175
- "\n",
1176
- " device = Config.DEVICE\n",
1177
- " model = load_mast3r_model(device)\n",
1178
- " scene, mast3r_images = run_mast3r_pairs(model, image_paths, pairs, device)\n",
1179
- "\n",
1180
- " del model\n",
1181
- " clear_memory()\n",
1182
- "\n",
1183
- " # STEP 4: Converting to COLMAP\n",
1184
- " print(\"\\n\" + \"=\"*70)\n",
1185
- " print(\"STEP 4: Converting to COLMAP (PINHOLE)\")\n",
1186
- " print(\"=\"*70)\n",
1187
- "\n",
1188
- " cameras_dict, pts3d, confidence = extract_camera_params_process2(\n",
1189
- " scene, image_paths, conf_threshold=conf_threshold\n",
1190
- " )\n",
1191
- "\n",
1192
- " del scene\n",
1193
- " clear_memory()\n",
1194
- "\n",
1195
- " if len(pts3d) > max_points:\n",
1196
- " print(f\"⚠️ Limiting points from {len(pts3d)} to {max_points}\")\n",
1197
- " indices = np.random.choice(len(pts3d), max_points, replace=False)\n",
1198
- " pts3d = pts3d[indices]\n",
1199
- " confidence = confidence[indices]\n",
1200
- "\n",
1201
- " print(f\"Final point count: {len(pts3d)}\")\n",
1202
- "\n",
1203
- " colmap_dir = os.path.join(output_dir, \"sparse/0\")\n",
1204
- " os.makedirs(colmap_dir, exist_ok=True)\n",
1205
- "\n",
1206
- " write_colmap_sparse(cameras_dict, pts3d, confidence, image_paths, colmap_dir)\n",
1207
- " clear_memory()\n",
1208
- "\n",
1209
- " # STEP 5: Running Gaussian Splatting\n",
1210
- " print(\"\\n\" + \"=\"*70)\n",
1211
- " print(\"STEP 5: Running Gaussian Splatting\")\n",
1212
- " print(\"=\"*70)\n",
1213
- "\n",
1214
- " source_dir = output_dir\n",
1215
- " model_output_dir = os.path.join(output_dir, \"gaussian_splatting\")\n",
1216
- "\n",
1217
- " gs_output = run_gaussian_splatting(\n",
1218
- " source_dir=source_dir,\n",
1219
- " output_dir=model_output_dir,\n",
1220
- " iterations=iterations\n",
1221
- " )\n",
1222
- "\n",
1223
- " # STEP 6: Verify Output\n",
1224
- " print(\"\\n\" + \"=\"*70)\n",
1225
- " print(\"PIPELINE COMPLETE\")\n",
1226
- " print(\"=\"*70)\n",
1227
- "\n",
1228
- " ply_path = os.path.join(\n",
1229
- " model_output_dir,\n",
1230
- " \"point_cloud\",\n",
1231
- " f\"iteration_{iterations}\",\n",
1232
- " \"point_cloud.ply\"\n",
1233
- " )\n",
1234
- "\n",
1235
- " if os.path.exists(ply_path):\n",
1236
- " file_size = os.path.getsize(ply_path) / (1024 * 1024)\n",
1237
- " print(f\"✓ Point cloud generated: {ply_path}\")\n",
1238
- " print(f\" Size: {file_size:.2f} MB\")\n",
1239
- " else:\n",
1240
- " print(f\"⚠️ Point cloud not found at: {ply_path}\")\n",
1241
- "\n",
1242
- " print(f\"\\nOutput directory structure:\")\n",
1243
- " print(f\" {output_dir}/\")\n",
1244
- " print(f\" ├── images/ (processed images)\")\n",
1245
- " if preprocess_mode == 'biplet':\n",
1246
- " print(f\" ├── original_images/ (original source images)\")\n",
1247
- " print(f\" ├── sparse/0/ (COLMAP data)\")\n",
1248
- " print(f\" └── gaussian_splatting/ (GS output)\")\n",
1249
- "\n",
1250
- " return gs_output\n",
1251
- "\n",
1252
- "# =====================================================================\n",
1253
- "# CELL 19: Verify Setup\n",
1254
- "# =====================================================================\n",
1255
- "print(f\"✓ np: {np.__version__} - {np.__file__}\")\n",
1256
- "!pip show numpy | grep Version\n",
1257
- "\n",
1258
- "try:\n",
1259
- " import roma\n",
1260
- " print(\"✓ roma is installed\")\n",
1261
- "except ModuleNotFoundError:\n",
1262
- " print(\"⚠️ roma not found, installing...\")\n",
1263
- " !pip install roma\n",
1264
- " import roma\n",
1265
- " print(\"✓ roma installed\")"
1266
- ],
1267
- "metadata": {
1268
- "trusted": true,
1269
- "id": "bHKT_3EZjLmH"
1270
- },
1271
- "outputs": [],
1272
- "execution_count": null
1273
- },
1274
- {
1275
- "cell_type": "code",
1276
- "source": [
1277
- "# =====================================================================\n",
1278
- "# CELL 20: Run Pipeline\n",
1279
- "# =====================================================================\n",
1280
- "if __name__ == \"__main__\":\n",
1281
- " IMAGE_DIR = \"/content/drive/MyDrive/your_folder/fountain\"\n",
1282
- " OUTPUT_DIR = \"/content/output\"\n",
1283
- "\n",
1284
- " gs_output = main_pipeline(\n",
1285
- " image_dir=IMAGE_DIR,\n",
1286
- " output_dir=OUTPUT_DIR,\n",
1287
- " square_size=800,\n",
1288
- " iterations=1000,\n",
1289
- " max_images=25,\n",
1290
- " max_pairs=25,\n",
1291
- " max_points=4000,\n",
1292
- " conf_threshold=1.5,\n",
1293
- " preprocess_mode='biplet'\n",
1294
- " )\n",
1295
- "\n",
1296
- " print(\"\\n\" + \"=\"*70)\n",
1297
- " print(\"PIPELINE COMPLETE\")\n",
1298
- " print(\"=\"*70)\n",
1299
- " print(f\"Output directory: {gs_output}\")"
1300
- ],
1301
- "metadata": {
1302
- "trusted": true,
1303
- "id": "n6ZHOb8TjLmI"
1304
- },
1305
- "outputs": [],
1306
- "execution_count": null
1307
- },
1308
- {
1309
- "cell_type": "code",
1310
- "source": [],
1311
- "metadata": {
1312
- "trusted": true,
1313
- "id": "Ontdbh48jLmI"
1314
- },
1315
- "outputs": [],
1316
- "execution_count": null
1317
- },
1318
- {
1319
- "cell_type": "markdown",
1320
- "source": [
1321
- "\n",
1322
- "\n",
1323
- "## 🔧 主要な修正:\n",
1324
- "\n",
1325
- "### 1. **特徴量抽出の修正 (CELL 12)**\n",
1326
- "- RGB画像 `[H, W, 3]` が返される問題を修正\n",
1327
- "- 特徴量次元が小さい場合は自動的に64次元に拡張\n",
1328
- "- より堅牢なエラーハンドリング\n",
1329
- "\n",
1330
- "### 2. **ASMK類似度計算の修正 (CELL 13)**\n",
1331
- "- Codebookの使用を削除し、シンプルなコサイン類似度に変更\n",
1332
- "- 次元ミスマッチエラーを完全に解消\n",
1333
- "- 動的な特徴量次元に対応\n",
1334
- "\n",
1335
- "### 3. **カメラパラメータの修正 (CELL 15)**\n",
1336
- "- 画像サイズ情報を明示的に保存 (`width`, `height`)\n",
1337
- "- より堅牢なエラーハンドリング\n",
1338
- "\n",
1339
- "### 4. **コード構造の改善**\n",
1340
- "- 各セルを独立して実行可能に\n",
1341
- "- メモリ管理の最適化\n",
1342
- "- エラーメッセージの改善\n",
1343
- "\n",
1344
- "## 📋 使用方法:\n",
1345
- "\n",
1346
- "1. **セル1**: 依存関係をインストール\n",
1347
- "2. **セル2**: カーネルを再起動（コメント）\n",
1348
- "3. **セル3-19**: 順番に実行\n",
1349
- "4. **セル20**: パイプラインを実行\n",
1350
- "\n",
1351
- "## ✨ 改善点:\n",
1352
- "\n",
1353
- "- ✅ ASMK失敗エラーを完全に解決\n",
1354
- "- ✅ 特徴量次元の動的対応\n",
1355
- "- ✅ メモリ効率の改善\n",
1356
- "- ✅ より詳細なログ出力\n",
1357
- "- ✅ エラー時の自動リカバリー\n",
1358
- "\n"
1359
- ],
1360
- "metadata": {
1361
- "id": "K-TGZRlcjLmI"
1362
- }
1363
- }
1364
- ]
1365
- }