feat: add action_state task description

action_state/gen_task1.py

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

action_state/gen_task2.py

@@ -0,0 +1,32 @@
+"""
+物体保持静止。摄像机围绕物体进行水平轨道运动。这里的“旋转”是指摄像机在水平面上绕物体中心移动，其方向定义参考鸟瞰视角下的时钟方向（即从上往下看，顺时针或逆时针移动）
+
+给定一张物体的初始视角图片，以及一系列旋转指令，模型需要计算这些指令叠加后的最终位置，并从A、B、C、D四个选项中选出该视角对应的图片。
+
+
+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 follows this sequence of rotations: {instruction_sequence}. Based on this top-down perspective, 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 according to the following instructions: {instruction_sequence}.
+
+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>
+
+"""

action_state/gen_task3.py

@@ -0,0 +1,57 @@
+"""
+**任务名称：** 围绕静止物体的相机视角排序（Camera View Ordering around Static Objects）
+
+**任务设定：**
+本任务模拟一个静止物体被移动相机拍摄的场景。相机围绕物体进行平滑的轨道旋转（Orbital Rotation）。
+1.  **数据来源：** 所有图像均从一段连续的视频中抽帧获得，因此具备内在的时空连贯性。
+2.  **运动约束：** 从给定的初始参考帧（Reference Frame）到序列的最后一帧，相机的总旋转角度严格限制在半个圆周（<180度）以内，保证视角的单向性和连续性。
+
+**输入数据：**
+*   一张**初始参考图（Reference Image）**，作为序列的起点（t=0）。
+*   四张**候选图片（Candidate Images）**，编号为 1、2、3、4。这四张图片是参考图之后的后续帧，但顺序已被打乱。
+
+**任务目标：**
+模型需要分析参考图与候选图片之间的几何关系，将四张候选图片按照真实的时间/空间顺序重新排列，使其构成一个连贯的相机运动轨迹。
+
+**输出格式：**
+提供 A、B、C、D 四个选项，每个选项代表一种排序组合（例如 B: 2-1-4-3）。模型需选择能够恢复正确时空顺序的选项。
+
+
+
+1.
+The object in the initial view <image_start>[image_ref]<image_end> remains **static**. Imagine a camera rotating around this object along a continuous path. The total rotation from the start to the end is less than 180 degrees.
+
+Below are four images captured during this rotation, labeled 1, 2, 3 and 4. They're currently shuffled.
+
+1. <image_start>[image_1]<image_end>
+2. <image_start>[image_2]<image_end>
+3. <image_start>[image_3]<image_end>
+4. <image_start>[image_4]<image_end>
+
+Please analyze the change in perspective and determine the correct chronological order of these four images following the initial view. Select the corret sequence.
+
+A. [sequence_A]
+B. [sequence_B]
+C. [sequence_C]
+D. [sequence_D]
+
+
+2.
+Given the initial view of the static object: <image_start>[image_ref]<image_end>.
+
+Imagine a camera rotating around this object to capture a video sequence. The rotation covers an anlge of less than 180 degrees. We have extracted four frames from the sequence, labeled 1 to 4, but their order is jumbled.
+
+1. <image_start>[image_1]<image_end>
+2. <image_start>[image_2]<image_end>
+3. <image_start>[image_3]<image_end>
+4. <image_start>[image_4]<image_end>
+
+Which of the following four options correctly sorts these images into a coherent spatio-temporal starting after the initial view?
+
+A. [sequence_A]
+B. [sequence_B]
+C. [sequence_C]
+D. [sequence_D]
+
+
+"""

action_state/gen_task4.py

@@ -0,0 +1,30 @@
+"""
+**任务名称：** 跨视角目标点匹配 / 旋转视角下的关键点对应
+
+**场景设定：**
+在一个静态场景中，保持物体位置固定，摄像机围绕物体进行旋转拍摄。
+
+**输入数据：**
+1.  **源视图（Source View）：** 摄像机在初始视角拍摄的物体图像，图像上标注了一个特定的关键点 $A$（Target Point）。
+2.  **目标视图（Target View）：** 摄像机旋转到新视角后拍摄的同一物体图像。该图像上标注了若干个候选点（例如：1, 2, 3, 4）。
+
+**任务目标：**
+模型需要分析两张图像的几何与纹理特征，从目标视图的候选点集合中，识别出哪一个点与源视图中的点 $A$ 对应于物体表面的同一物理位置（例如，如果点 2 与点 $A$ 是同一点，模型应输出 2）。
+
+
+1.
+The object in the initial view <image_start>[image_1]<image_end> remains **static**, with a specific point marked as **A**.
+
+Imagine a camera rotating around this object to a new position. The final view captured from this new angle is shown below, with four candidate points marked as **1, 2, 3 and 4**.
+
+Based on the visual features and geometry, which of the numbered points in the final view corresponds to the **same physical location** as point **A** in the inital view?
+
+**Final view**：<image_start>[image_2]<image_end>
+
+A. Point 1
+B. Point 2
+C. Point 3
+D. Point 4
+
+
+"""

action_state/utils.py

@@ -0,0 +1,87 @@
+import json
+import os
+import numpy as np
+from scipy.spatial.transform import Rotation as SciRotation
+
+def get_relative_horizontal_rotation(root_path, category, sequence_number, frame_idx_1, frame_idx_2):
+    """
+    计算 frame_2 相对于 frame_1 在水平方向上的旋转角度 (Yaw)。
+    
+    Args:
+        root_path (str): 数据集根目录 (包含 data/original/...)
+        category (str): 类别名称 (例如 'motorcycle')
+        sequence_number (str): 序列号 (例如 '613_98146_195503')
+        frame_idx_1 (int): 起始帧号 (例如 6)
+        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。
+               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')
+
+    # 2. 读取 JSON 数据
+    with open(json_path, 'r') as f:
+        annotations = json.load(f)
+
+    # 3. 查找对应帧的 R 矩阵
+    R1 = None
+    R2 = None
+    
+    # 这里的 annotations 可能是一个大列表，包含该 category 下所有 sequence 的帧
+    # 或者如果 json 在 sequence 目录下，它只包含该 sequence 的帧
+    for item in annotations:
+        # 过滤 sequence (如果 json 包含多个 sequence)
+        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:
+        print(f"Error: Could not find frame {frame_idx_1} or {frame_idx_2} in sequence {sequence_number}")
+        return None
+
+    # 4. 计算相对旋转
+    # 假设 R 是 World-to-Camera 矩阵 (常见于 NeRF/PyTorch3D 数据)
+    # 我们想要计算 Camera 1 到 Camera 2 的变换。
+    # R_rel = R2 @ R1.T
+    # 解释: R1.T 将 Cam1 转回 World，R2 将 World 转去 Cam2
+    R_rel = R2 @ R1.T
+    
+    # 5. 转换为欧拉角提取水平分量
+    # 使用 scipy 的 Rotation 库处理，避免万向节死锁和复杂的数学公式
+    r = SciRotation.from_matrix(R_rel)
+    
+    # 这里的顺序 'y' 代表水平轴旋转 (Yaw)。
+    # 常见的顺序是 'xyz', 'zyx' 等。
+    # 对于大多数 3D 视觉任务（Y-up 坐标系），Y 轴旋转即为水平旋转。
+    # 如果是 Z-up 坐标系，则需要取 'z' 分量。
+    # CO3D / PyTorch3D 通常是 Y-up, X-left, Z-in (NDC) 或者类似的右手系。
+    # 我们提取 'y' 分量作为水平角。
+    
+    # as_euler 返回的是 [x_angle, y_angle, z_angle] (对应 'xyz' 顺序)
+    # degrees=True 直接返回角度
+    euler_angles = r.as_euler('xyz', degrees=True) 
+    
+    # 提取 Y 轴分量 (索引 1)
+    horizontal_rotation = euler_angles[1]
+    
+    return horizontal_rotation
+
+# --- 测试代码 ---
+if __name__ == "__main__":
+    yaw = get_relative_horizontal_rotation(
+        root_path="/run/determined/NAS1/public/lixinyuan/interleaved-co3d",
+        category="motorcycle",
+        sequence_number="216_22798_47409",
+        frame_idx_1=90,
+        frame_idx_2=74,
+    )
+    
+    print(f"Calculated Relative Yaw: {yaw:.4f} degrees")