File size: 16,857 Bytes
c7f75e8 b73b7d0 c7f75e8 b73b7d0 c7f75e8 b73b7d0 c7f75e8 b73b7d0 c7f75e8 b73b7d0 c7f75e8 a104ac4 24d1e42 3e0a7cf c00979a a104ac4 c00979a 24d1e42 8e74163 60fde3b c00979a 24d1e42 b73b7d0 24d1e42 c00979a 60fde3b c00979a 60fde3b c7f75e8 8e74163 a104ac4 24d1e42 a104ac4 8e74163 b73b7d0 24d1e42 a104ac4 8e74163 b73b7d0 24d1e42 a104ac4 60fde3b c00979a 60fde3b c00979a 60fde3b c00979a 24d1e42 c00979a 24d1e42 8e74163 60fde3b 8e74163 24d1e42 60fde3b a104ac4 24d1e42 c00979a a104ac4 60fde3b a104ac4 24d1e42 a104ac4 60fde3b 24d1e42 c00979a 24d1e42 8e74163 24d1e42 8e74163 c7f75e8 24d1e42 a104ac4 c00979a 60fde3b c00979a 24d1e42 c00979a a104ac4 24d1e42 a104ac4 24d1e42 c00979a c7f75e8 24d1e42 c7f75e8 24d1e42 a104ac4 24d1e42 a104ac4 c7f75e8 24d1e42 c7f75e8 60fde3b 24d1e42 c7f75e8 24d1e42 a104ac4 60fde3b 24d1e42 60fde3b a104ac4 24d1e42 c00979a 60fde3b c00979a 60fde3b c00979a 60fde3b c00979a 60fde3b c00979a 60fde3b c00979a 60fde3b c00979a c7f75e8 24d1e42 c7f75e8 24d1e42 c7f75e8 24d1e42 c7f75e8 24d1e42 c7f75e8 24d1e42 c7f75e8 60fde3b a104ac4 24d1e42 a104ac4 60fde3b a104ac4 c7f75e8 60fde3b c7f75e8 60fde3b c7f75e8 a104ac4 24d1e42 a104ac4 24d1e42 b73b7d0 c00979a a104ac4 24d1e42 a104ac4 24d1e42 b73b7d0 a104ac4 24d1e42 b73b7d0 a104ac4 24d1e42 a104ac4 24d1e42 a104ac4 24d1e42 a104ac4 24d1e42 a104ac4 24d1e42 a104ac4 c7f75e8 24d1e42 3e0a7cf c00979a 3e0a7cf c00979a 3e0a7cf c00979a 3e0a7cf c00979a 3e0a7cf c7f75e8 24d1e42 c7f75e8 a104ac4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 | """
给定一张物体在初始视角的图片,camera将围绕该静止物体进行水平移动。旋转的方向(顺时针或逆时针)是基于从物体正上方俯视(鸟瞰视角)的平面来定义的。模型需要根据给定的旋转方向和角度,推断出新视角下的物体图像,并从四个图像中选出正确的一项。
注:{angle}就是 xx degrees, {direction} 就是 clockwise / anticlockwise
1.
The {object} in the image <image_start>[image_1]<image_end> remains **static**. Imagine a camera rotating around this {object}. The direction of rotation is defined from a **top-down bird's-eye view**.
Please identify the view of the {object} after the camera rotates {angle} {direction} based on this top-down perspective, and select the correct answer.
A. <image_start>[image_A]<image_end>
B. <image_start>[image_B]<image_end>
C. <image_start>[image_C]<image_end>
D. <image_start>[image_D]<image_end>
2.
Given the initial view of a **static** {object}: <image_start>[image_1]<image_end>.
Imagine looking at the setup from a bird's-eye view (from directly above) to determine the direction. Now, move the camera {angle} {direction} around the {object}.
Which of the following images shows what the {object} looks like from this new position?
A. <image_start>[image_A]<image_end>
B. <image_start>[image_B]<image_end>
C. <image_start>[image_C]<image_end>
D. <image_start>[image_D]<image_end>
"""
import argparse
import random
import json
import os
from tqdm import tqdm
import numpy as np # 确保导入 numpy
# 导入公共工具库
from utils import (
CO3DDataLoader,
get_relative_yaw,
format_angle_direction,
get_angle_diff,
format_image_path,
save_jsonl_splits,
get_sequence_geometry_pca,
decompose_angle
)
class Task1Generator:
def __init__(self, loader, image_prefix):
self.loader = loader
self.image_prefix = image_prefix
self.cat_name = self.loader.category.replace('_', ' ')
# 定义旋转步长配置,与 create_entry 保持一致
self.ROTATION_STEPS = [30, 15]
# 定义允许的最大误差(度):如果模拟角度和真实图片角度相差超过此值,则认为该样本无效
self.MAX_COT_ERROR = 10.0
def verify(self, start_R, start_T, target_R, target_T, distractor_infos,
min_angle, max_angle, min_interval, mean_center, basis):
"""
Task 1 专用验证逻辑:
确保 Start, Target, Distractors 任意两张图之间的角度差都大于 min_interval
"""
target_yaw = get_relative_yaw(start_R, start_T, target_R, target_T, mean_center, basis)
if not (min_angle <= abs(target_yaw) <= max_angle):
return False, None, []
distractor_yaws = []
for d_info in distractor_infos:
d_yaw = get_relative_yaw(start_R, start_T, d_info['R'], d_info['T'], mean_center, basis)
distractor_yaws.append(d_yaw)
all_angles = [0.0, target_yaw] + distractor_yaws
for i in range(len(all_angles)):
for j in range(i + 1, len(all_angles)):
if get_angle_diff(all_angles[i], all_angles[j]) < min_interval:
return False, None, []
return True, target_yaw, distractor_yaws
def _get_all_relative_angles(self, start_idx, all_frames, seq_data_dict, mean_center, basis):
"""
计算序列中所有帧相对于 start_idx 的角度。
"""
start_info = seq_data_dict[start_idx]
results = []
for f_idx in all_frames:
if f_idx == start_idx:
results.append({'idx': f_idx, 'angle': 0.0})
continue
f_info = seq_data_dict[f_idx]
yaw = get_relative_yaw(
start_info['R'], start_info['T'],
f_info['R'], f_info['T'],
mean_center, basis
)
results.append({'idx': f_idx, 'angle': yaw})
return results
def _check_cot_feasibility(self, target_yaw, all_rel_data):
"""
[新增] 检查 CoT 路径的可行性。
如果中间某一步找不到足够接近的真实图片(误差 > MAX_COT_ERROR),则返回 False。
"""
rotation_sign = 1 if target_yaw >= 0 else -1
total_delta = abs(target_yaw)
steps = decompose_angle(total_delta, self.ROTATION_STEPS)
current_simulated_angle = 0.0
for step in steps:
current_simulated_angle += (step * rotation_sign)
# 寻找最近邻的角度差
min_diff = float('inf')
for item in all_rel_data:
diff = abs(item['angle'] - current_simulated_angle)
if diff < min_diff:
min_diff = diff
# 如果最近的一张图误差都很大,说明这里缺帧,不能生成高质量 CoT
if min_diff > self.MAX_COT_ERROR:
return False
return True
def generate_sample(self, seq_name, config):
frames = self.loader.get_frames(seq_name)
if len(frames) < 10: # 稍微提高一点门槛,太短的序列很难凑齐中间帧
return None
seq_data_dict = self.loader.seq_data[seq_name]
mean_center, basis, _ = get_sequence_geometry_pca(seq_data_dict)
max_attempts = 5000
for _ in range(max_attempts):
# A. 随机采样
start_idx = random.choice(frames)
start_info = seq_data_dict[start_idx]
possible_targets = [f for f in frames if f != start_idx]
if not possible_targets: continue
target_idx = random.choice(possible_targets)
target_info = seq_data_dict[target_idx]
remaining = [f for f in frames if f != start_idx and f != target_idx]
if len(remaining) < 3: continue
distractor_indices = random.sample(remaining, 3)
distractor_infos = [seq_data_dict[d] for d in distractor_indices]
# B. 验证几何约束 (Start/Target/Distractors 之间的互斥性)
is_valid, target_yaw, distractor_yaws = self.verify(
start_info['R'], start_info['T'],
target_info['R'], target_info['T'],
distractor_infos,
config['min_angle'],
config['max_angle'],
config['min_interval'],
mean_center, basis
)
if is_valid:
# === 关键修改:先获取所有帧角度,进行 CoT 可行性预检查 ===
all_rel_data = self._get_all_relative_angles(
start_idx, frames, seq_data_dict, mean_center, basis
)
# 如果 CoT 路径中间缺图,直接跳过,重新采样
if not self._check_cot_feasibility(target_yaw, all_rel_data):
continue
return self.create_entry(
seq_name, start_idx, target_idx, distractor_indices,
target_yaw, distractor_yaws, start_info, target_info, distractor_infos,
all_rel_data
)
return None
def create_entry(self, seq_name, start_idx, target_idx, distractor_indices,
target_yaw, distractor_yaws, start_info, target_info, distractor_infos,
all_rel_data):
angle_deg, direction_str = format_angle_direction(target_yaw)
# 1. 构建选项列表
options = [{
"path": format_image_path(target_info['path'], self.loader.root_path, self.image_prefix),
"angle": target_yaw,
"is_correct": True
}]
for d_idx, d_yaw, d_info in zip(distractor_indices, distractor_yaws, distractor_infos):
options.append({
"path": format_image_path(d_info['path'], self.loader.root_path, self.image_prefix),
"angle": d_yaw,
"is_correct": False
})
random.shuffle(options)
# 2. 映射到 A, B, C, D (保持不变,但记录正确答案的 Label)
images_dict = {
"image_1": format_image_path(start_info['path'], self.loader.root_path, self.image_prefix)
}
option_labels = ['A', 'B', 'C', 'D']
correct_label = ""
# 用于 Oracle 数据的选项信息
options_meta = {}
for label, opt in zip(option_labels, options):
img_key = f"image_{label}"
images_dict[img_key] = opt["path"]
options_meta[label] = {
"image_key": img_key,
"angle": opt["angle"],
"is_correct": opt["is_correct"]
}
if opt["is_correct"]:
correct_label = label
# ------------------------------------------------------------------
# 3. 构建 Oracle Chain (核心修改)
# ------------------------------------------------------------------
rotation_sign = 1 if target_yaw >= 0 else -1
total_delta = abs(target_yaw)
steps = decompose_angle(total_delta, self.ROTATION_STEPS)
oracle_chain = []
current_simulated_angle = 0.0
# 记录每一步的详细信息
for step_idx, step_angle in enumerate(steps):
current_simulated_angle += (step_angle * rotation_sign)
# 寻找最近邻帧
closest_frame_data = min(all_rel_data, key=lambda x: abs(x['angle'] - current_simulated_angle))
closest_frame_idx = closest_frame_data['idx']
closest_frame_info = self.loader.get_frame_info(seq_name, closest_frame_idx)
# 定义这一步产生的中间图的 key
reasoning_key = f"reasoning_image_{step_idx + 1}"
images_dict[reasoning_key] = format_image_path(
closest_frame_info['path'], self.loader.root_path, self.image_prefix
)
# 构造 Chain Item
chain_item = {
"step_index": step_idx + 1,
"action": {
"type": "rotate",
"degrees": step_angle,
"direction": direction_str,
"total_angle_so_far": current_simulated_angle
},
"result_image_key": reasoning_key,
"is_final_step": (step_idx == len(steps) - 1)
}
oracle_chain.append(chain_item)
# ------------------------------------------------------------------
# 4. 生成 Prompt
# ------------------------------------------------------------------
template_id = random.choice([1, 2])
if template_id == 1:
question = f"""The {self.cat_name} in the image <image_start>[image_1]<image_end> remains **static**. Imagine a camera rotating around this {self.cat_name}. The direction of rotation is defined from a **top-down bird's-eye view**.
Please identify the view of the {self.cat_name} after the camera rotates {angle_deg} degrees {direction_str} based on this top-down perspective, and select the correct answer.
A. <image_start>[image_A]<image_end>
B. <image_start>[image_B]<image_end>
C. <image_start>[image_C]<image_end>
D. <image_start>[image_D]<image_end>"""
else:
question = f"""Given the initial view of a **static** {self.cat_name}: <image_start>[image_1]<image_end>.
Imagine looking at the setup from a bird's-eye view (from directly above) to determine the direction. Now, move the camera {angle_deg} degrees {direction_str} around the {self.cat_name}.
Which of the following images shows what the {self.cat_name} looks like from this new position?
A. <image_start>[image_A]<image_end>
B. <image_start>[image_B]<image_end>
C. <image_start>[image_C]<image_end>
D. <image_start>[image_D]<image_end>"""
# ------------------------------------------------------------------
# 5. 返回结构化数据 (不再包含 cot_trace 字符串)
# ------------------------------------------------------------------
return {
"id": f"task1_{seq_name}_{start_idx}_{target_idx}",
"task": "camera_view_prediction",
"sequence": seq_name,
"question": question,
"images": images_dict, # 包含 start, options, 和所有的 reasoning_images
"oracle_meta": {
"start_frame": start_idx,
"target_frame": target_idx,
"angle_degrees": angle_deg,
"direction": direction_str,
"correct_label": correct_label,
"options": options_meta,
"chain": oracle_chain # 这是给 LLM 看的“剧本”
},
# gt_answer 字段暂时留空或只放答案,等 Stage 2 再填入完整的 CoT
"gt_answer": f"<answer>{correct_label}</answer>"
}
def main():
parser = argparse.ArgumentParser(description="Generate Task 1: Camera View Prediction")
# 路径配置
parser.add_argument("--root_path", type=str, required=True, help="CO3D dataset root")
parser.add_argument("--output_dir", type=str, default="output_task1", help="Output directory")
parser.add_argument("--image_prefix", type=str, default="data/", help="Prefix for image paths")
parser.add_argument("--filter_path", type=str, default=None, help="Root directory for filter logs")
# 采样配置
parser.add_argument("--category", type=str, default=None, help="Specific category or None for all")
parser.add_argument("--num_samples", type=int, default=1, help="Samples per sequence")
parser.add_argument("--seed", type=int, default=42)
# 几何约束配置
parser.add_argument("--min_angle", type=float, default=40.0)
parser.add_argument("--max_angle", type=float, default=140.0)
parser.add_argument("--min_interval", type=float, default=25.0)
# 切分配置
parser.add_argument("--train_ratio", type=float, default=0.8)
parser.add_argument("--val_ratio", type=float, default=0.1)
parser.add_argument("--test_ratio", type=float, default=0.1)
parser.add_argument("--max_items", type=int, default=10000)
args = parser.parse_args()
random.seed(args.seed)
np.random.seed(args.seed)
if args.category:
categories = [args.category]
else:
data_dir = os.path.join(args.root_path, 'data', 'original')
if os.path.exists(data_dir):
categories = sorted([d for d in os.listdir(data_dir) if os.path.isdir(os.path.join(data_dir, d))])
else:
print(f"Error: {data_dir} not found.")
return
all_results = []
config = {
'min_angle': args.min_angle,
'max_angle': args.max_angle,
'min_interval': args.min_interval
}
for cat in categories:
loader = CO3DDataLoader(args.root_path, cat)
if not loader.seq_data:
continue
generator = Task1Generator(loader, args.image_prefix)
sequences = loader.get_sequences()
if args.filter_path:
keep_file = os.path.join(args.filter_path, cat, "keep.json")
if os.path.exists(keep_file):
try:
with open(keep_file, 'r') as f:
keep_list = set(json.load(f))
sequences = [s for s in sequences if s in keep_list]
print(f"[{cat}] Filter applied: {len(sequences)} sequences retained.")
except Exception as e:
print(f"[{cat}] Error reading keep.json: {e}. Skipping.")
sequences = []
else:
print(f"[{cat}] Warning: No keep.json found. Skipping.")
sequences = []
if not sequences:
continue
for seq in tqdm(sequences, desc=f"Task1 - {cat}", leave=False):
for _ in range(args.num_samples):
sample = generator.generate_sample(seq, config)
if sample:
all_results.append(sample)
print(f"Total generated: {len(all_results)}")
save_jsonl_splits(
all_results,
args.output_dir,
ratios=(args.train_ratio, args.val_ratio, args.test_ratio),
max_items=args.max_items,
seed=args.seed
)
print(f"Done. Output saved to {args.output_dir}")
if __name__ == "__main__":
main()
|