refactor

action_state/gen_task1.py

@@ -1,29 +0,0 @@
-"""
-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>
-
-"""

src/action_state/gen_task1.py

@@ -0,0 +1,275 @@
+"""
+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 json
+import os
+import random
+import numpy as np
+from scipy.spatial.transform import Rotation as SciRotation
+
+# --- 1. 引入你提供的 Utils 函数 ---
+
+def get_relative_horizontal_rotation(root_path, category, sequence_number, frame_idx_1, frame_idx_2, annotations_cache=None):
+    """
+    计算 frame_2 相对于 frame_1 在水平方向上的旋转角度 (Yaw)。
+    增加了 annotations_cache 以避免重复读取大文件。
+    """
+    if annotations_cache is None:
+        json_path = os.path.join(root_path, 'data', 'original', category, 'frame_annotations.json')
+        with open(json_path, 'r') as f:
+            annotations = json.load(f)
+    else:
+        annotations = annotations_cache
+
+    R1 = None
+    R2 = None
+    
+    for item in annotations:
+        if item.get('sequence_name') != sequence_number:
+            continue
+        if item['frame_number'] == frame_idx_1:
+            R1 = np.array(item['viewpoint']['R'])
+        elif item['frame_number'] == frame_idx_2:
+            R2 = np.array(item['viewpoint']['R'])
+        if R1 is not None and R2 is not None:
+            break
+    
+    if R1 is None or R2 is None:
+        return None
+
+    # 计算相对旋转 R_rel = R2 @ R1.T
+    R_rel = R2 @ R1.T
+    
+    # 转换为欧拉角提取水平分量 (Y轴)
+    r = SciRotation.from_matrix(R_rel)
+    euler_angles = r.as_euler('xyz', degrees=True) 
+    horizontal_rotation = euler_angles[1]
+    
+    return horizontal_rotation
+
+# --- 2. 核心生成逻辑 ---
+
+class CO3DQuestionGenerator:
+    def __init__(self, root_path, category, output_file="co3d_rotation_questions.json"):
+        self.root_path = root_path
+        self.category = category
+        self.output_file = output_file
+        
+        # 预加载 annotations 以提高速度
+        print(f"Loading annotations for category: {self.category}...")
+        json_path = os.path.join(root_path, 'data', 'original', category, 'frame_annotations.json')
+        with open(json_path, 'r') as f:
+            self.annotations = json.load(f)
+        
+        # 整理数据结构：Sequence -> List of Frames
+        self.seq_map = {}
+        for item in self.annotations:
+            seq_name = item['sequence_name']
+            if seq_name not in self.seq_map:
+                self.seq_map[seq_name] = []
+            self.seq_map[seq_name].append(item['frame_number'])
+        
+        # 确保帧是排序的，虽然对于随机采样不是必须的，但有助于调试
+        for seq in self.seq_map:
+            self.seq_map[seq].sort()
+
+    def format_angle_direction(self, angle):
+        """
+        将带符号的角度转换为绝对值角度和方向字符串。
+        假设：正值 = 顺时针 (Clockwise), 负值 = 逆时针 (Anticlockwise)
+        注意：具体正负定义取决于坐标系，这里基于 utils 中的逻辑假设。
+        """
+        # 归一化到 -180 到 180
+        angle = (angle + 180) % 360 - 180
+        
+        direction = "clockwise" if angle > 0 else "anticlockwise"
+        abs_angle = abs(angle)
+        
+        return int(round(abs_angle)), direction
+
+    def get_image_path(self, sequence_name, frame_idx):
+        # 假设图片路径结构，根据实际情况调整
+        # 例如: data/original/motorcycle/sequence_name/frame00001.jpg
+        # 这里假设 frame_idx 是整数，文件名可能是 0 填充的
+        # 注意：CO3D 的文件名通常在 annotation 的 'image' 字段里，这里为了简化直接拼凑，
+        # 如果文件名不规则，建议建立 frame_idx -> image_path 的映射。
+        
+        # 更稳健的方法是从 annotations 找 image path
+        for item in self.annotations:
+            if item['sequence_name'] == sequence_name and item['frame_number'] == frame_idx:
+                # item['image']['path'] 通常是相对路径
+                return os.path.join(self.root_path, item['image']['path'])
+        return None
+
+    def generate_sample(self, sequence_name):
+        frames = self.seq_map[sequence_name]
+        if len(frames) < 5: # 至少需要 1个起始 + 1个正确 + 3个干扰
+            return None
+
+        # 1. 随机选择起始帧 (Start Frame)
+        start_frame_idx = random.choice(frames)
+        
+        # 2. 随机选择目标帧 (Target Frame / Correct Answer)
+        # 为了保证题目有意义，角度最好不要太小（比如 < 10度）
+        possible_targets = [f for f in frames if f != start_frame_idx]
+        random.shuffle(possible_targets)
+        
+        target_frame_idx = None
+        angle_val = 0
+        
+        for t_idx in possible_targets:
+            yaw = get_relative_horizontal_rotation(
+                self.root_path, self.category, sequence_name, 
+                start_frame_idx, t_idx, self.annotations
+            )
+            if yaw is None: continue
+            
+            # 过滤掉角度过小的样本 (例如小于 15 度)
+            if abs(yaw) > 15:
+                target_frame_idx = t_idx
+                angle_val = yaw
+                break
+        
+        if target_frame_idx is None:
+            return None
+
+        # 3. 选择干扰项 (Distractors)
+        # 干扰项不能是起始帧，也不能是目标帧
+        remaining_frames = [f for f in frames if f != start_frame_idx and f != target_frame_idx]
+        if len(remaining_frames) < 3:
+            return None
+        
+        distractor_indices = random.sample(remaining_frames, 3)
+        
+        # 4. 准备数据
+        angle_deg, direction_str = self.format_angle_direction(angle_val)
+        
+        # 获取图片路径
+        img_start = self.get_image_path(sequence_name, start_frame_idx)
+        img_correct = self.get_image_path(sequence_name, target_frame_idx)
+        img_distractors = [self.get_image_path(sequence_name, idx) for idx in distractor_indices]
+        
+        if not all([img_start, img_correct] + img_distractors):
+            return None
+
+        # 5. 构建选项
+        options = [img_correct] + img_distractors
+        random.shuffle(options) # 打乱选项顺序
+        
+        correct_option_idx = options.index(img_correct)
+        option_labels = ['A', 'B', 'C', 'D']
+        correct_answer_label = option_labels[correct_option_idx]
+
+        # 6. 构建 Prompt (随机选择模板 1 或 2)
+        template_id = random.choice([1, 2])
+        
+        prompt = ""
+        if template_id == 1:
+            prompt = f"""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_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:
+            prompt = f"""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_deg} degrees {direction_str} 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>"""
+
+        # 构建返回对象
+        sample = {
+            "id": f"{sequence_name}_{start_frame_idx}_{target_frame_idx}",
+            "sequence": sequence_name,
+            "category": self.category,
+            "prompt": prompt,
+            "images": {
+                "image_1": img_start,
+                "image_A": options[0],
+                "image_B": options[1],
+                "image_C": options[2],
+                "image_D": options[3]
+            },
+            "metadata": {
+                "start_frame": start_frame_idx,
+                "target_frame": target_frame_idx,
+                "angle_degrees": angle_deg,
+                "direction": direction_str,
+                "raw_yaw": angle_val
+            },
+            "answer": correct_answer_label
+        }
+        return sample
+
+    def run(self, num_samples=100):
+        results = []
+        sequences = list(self.seq_map.keys())
+        
+        count = 0
+        attempts = 0
+        max_attempts = num_samples * 5 # 防止死循环
+        
+        print(f"Start generating {num_samples} samples...")
+        
+        while count < num_samples and attempts < max_attempts:
+            attempts += 1
+            seq = random.choice(sequences)
+            sample = self.generate_sample(seq)
+            
+            if sample:
+                results.append(sample)
+                count += 1
+                if count % 10 == 0:
+                    print(f"Generated {count}/{num_samples} samples...")
+        
+        print(f"Finished. Saving to {self.output_file}")
+        with open(self.output_file, 'w') as f:
+            json.dump(results, f, indent=4)
+
+# --- 3. 运行脚本 ---
+
+if __name__ == "__main__":
+    # 配置参数
+    ROOT_PATH = "/run/determined/NAS1/public/lixinyuan/interleaved-co3d" # 你的数据根目录
+    CATEGORY = "motorcycle" # 你的类别
+    OUTPUT_JSON = "co3d_motorcycle_questions.json"
+    NUM_SAMPLES = 50 # 你想生成的题目数量
+
+    generator = CO3DQuestionGenerator(ROOT_PATH, CATEGORY, OUTPUT_JSON)
+    generator.run(NUM_SAMPLES)

{action_state → src/action_state}/utils.py

@@ -15,7 +15,7 @@ def get_relative_horizontal_rotation(root_path, category, sequence_number, frame
         frame_idx_2 (int): 目标帧号 (例如 7)
         
     Returns:
-        float: 相对旋转角度（单位：度 Degree），正值通常表示向右/逆时针旋转，取决于坐标系定义。
+        float: 相对旋转角度（单位：度 Degree），正值表示顺时针旋转。
                如果找不到文件或帧，返回 None。
     """
     json_path = os.path.join(root_path, 'data', 'original', category, 'frame_annotations.json')