evaluation_00001024_camera

直接读取 v1.1 bundle（evaluation_00001024 的 exo 视图，split=test，共 512 帧独立图像）。预览仅供人工复核，不是训练输入（CONTRACT §6.5）；本次只解码抽样的 6 帧，不是全部帧。本 release 另带 v1.0 的 24 张抽查页（doc/history-v1.0/validation/），两者并存，本页不替代它。

配色沿用 RHD v1.0 复核页自己的约定（「红左手、蓝右手」），便于与 ../history-v1.0/validation/ 里的 24 张抽查页并排读：    左手 ·     右手
手槽取自发布的 hand_slots: [left, right]，21 个关节保持发布的近端→远端顺序（keypoint_schema: rhd21/v1，原生 tip→palm 已按 source_index_map 重排）。RHD 的 21 点只统一其原生体系顺序，不宣称与 MANO 关节位置等价；release 里没有 MANO、没有物体、没有接触，这些面板也不画。

本页能做别人做不到的一件事：RHD 同时发布了原生 2D 关键点（uv）和相机系 3D 关键点（xyz）以及同一套针孔 K，所以第 4 个面板可以把 xyz 经该视图自己的 K 投到像素上，再与发布的 uv 逐点比对（H2O/dex-ycb/ARCTIC 都缺其中一条序列，做不了这项对照）。偏差记在每帧的projection.frames 里，面板标题上也会写出中位与最大偏离；偏差只被记录，不用来「修正」发布值。

全视图数值扫描（不止抽样帧，本视图 512 帧全部参与）：发布的 xyz 经本视图自己的 K 投影后与发布的 uv 逐点比对。画内（发布的 uv 落在 320×320 内）15642 个点：中位偏离 0.0382 px、最大 0.1224 px；512/512 帧的画内点全部在 1 px 以内。画外 5527 个点：发布的 uv 是原始投影，可以任意大，像素阈值对它没有意义，故按相对自身量级核验——其中 5458 个点绝对偏差仍在 1 px 以内。两个极值来自不同的点，故各自带上它自己的量级：绝对偏差最大 41.5559 px（该点坐标量级 1.51e+05 px，相对 0.00027）；相对偏差最大 0.0031（该点绝对偏差仅 0.0741 px、离原点 23.9 px，比值大只因分母小）。另有 331 个关节深度 ≤0（在相机平面上或之后）、4 个深度 < 0.001 m（其 uv 本身就在 1e5 px 量级），这两类不成对比较，只计数。偏差最大的画内帧已强制进抽样，见下方。

RHD 每张图标注双手，而通常只有一只手落在 320×320 画内（本视图 15642/21504 个关节落在画内），源用 visibility 标出每个关节是否可见。实测该标量与「发布的 uv 是否落在画内」在绝大多点上一致，不一致的全部是「画内但被标为不可见」（本视图 461 个，多为被另一只手遮挡），而「标为可见却落在画外」只有 0 个。发布的 uv 是原始投影值，画外关节的坐标可以任意大；本页按发布的 uv 是否落在画内分两组核验——画内点按像素偏差、画外点按相对自身量级的偏差，另单列"源标为可见却落在画外"的点数（该方向才可能是标注与标定不一致）。可见与不可见只影响绘制时的实心/空心，不影响有效性判定：不可见关节仍然有源给出的 xyz/uv，故仍参与上述比对。画外关节占多数是本数据集「每图标注双手、通常只有一只在画内」的正常形态，不是标定缺陷。

帧 0 / 512

原生 RGB（MKV 解出的第 0 帧，未重编码）
分割区域（左手红 3058 px / 右手蓝 960 px，叠加在 RGB 上；区域号 2–33，背景号 65534）
原生 UV 关键点（发布的 2D 像素坐标；实心＝源标为可见（基本等于落在画内）、空心＝源标为不可见（画外或被遮挡）；左手 21 点、右手 9 点在画内）
XYZ 投影（相机系米制经本视图自己的 K；左手 21 点、右手 9 点在画内，实心＝源标为可见）；与发布 uv 逐点比对（画内 30 点）：中位偏离 0.0303 px、最大 0.0614 px，≤1 px 的 30 点；画外 12 点按自身量级的最大相对偏离 0.000341
z 深度（float32 位串，逐图拉伸；显示范围 0.496–5.000 m）
帧号、来源、全部数值组件与有效性
{
  "frame_index": 0,
  "source": {
    "frame_index": 0,
    "source_frame_id": "1024",
    "timestamp_ns": null,
    "clock_domain": null,
    "source_locator": {
      "archive": "raw/data/RHD_v1-1.zip",
      "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle",
      "source_frame_id": "1024",
      "annotation_index": 1024,
      "member": "RHD_published_v2/evaluation/color/01024.png"
    },
    "sample_id": "RHD/evaluation/01024"
  },
  "components": {
    "intrinsic": {
      "shape": [
        512,
        3,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 512,
      "valid_total": 512,
      "nan_count": 0
    },
    "keypoints_xyz": {
      "shape": [
        512,
        2,
        21,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_xyz.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    },
    "keypoints_uv": {
      "shape": [
        512,
        2,
        21,
        2
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_uv.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    }
  },
  "media_valid": {
    "rgb": true,
    "depth_true_pixels": 102400,
    "depth_pixels": 102400,
    "depth_all_true": true
  },
  "projection": {
    "annotated": true,
    "valid_points": 42,
    "behind_plane_points": 0,
    "near_plane_points": 0,
    "in_frame_points": 30,
    "outside_frame_points": 12,
    "visible_outside_frame": 0,
    "uv_compared": 30,
    "uv_median_px": 0.0303,
    "uv_max_px": 0.0614,
    "uv_within_1px": 30,
    "uv_within_tolerance": 30,
    "outside_uv_max_px": 0.028,
    "outside_uv_max_relative": 0.000341
  }
}

帧 128 / 512

原生 RGB（MKV 解出的第 128 帧，未重编码）
分割区域（左手红 3126 px / 右手蓝 0 px，叠加在 RGB 上；区域号 2–33，背景号 65534）
原生 UV 关键点（发布的 2D 像素坐标；实心＝源标为可见（基本等于落在画内）、空心＝源标为不可见（画外或被遮挡）；左手 21 点、右手 0 点在画内）
XYZ 投影（相机系米制经本视图自己的 K；左手 21 点、右手 0 点在画内，实心＝源标为可见）；与发布 uv 逐点比对（画内 21 点）：中位偏离 0.0321 px、最大 0.0553 px，≤1 px 的 21 点；画外 21 点按自身量级的最大相对偏离 0.000424
z 深度（float32 位串，逐图拉伸；显示范围 0.463–5.000 m）
帧号、来源、全部数值组件与有效性
{
  "frame_index": 128,
  "source": {
    "frame_index": 128,
    "source_frame_id": "1152",
    "timestamp_ns": null,
    "clock_domain": null,
    "source_locator": {
      "archive": "raw/data/RHD_v1-1.zip",
      "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle",
      "source_frame_id": "1152",
      "annotation_index": 1152,
      "member": "RHD_published_v2/evaluation/color/01152.png"
    },
    "sample_id": "RHD/evaluation/01152"
  },
  "components": {
    "intrinsic": {
      "shape": [
        512,
        3,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 512,
      "valid_total": 512,
      "nan_count": 0
    },
    "keypoints_xyz": {
      "shape": [
        512,
        2,
        21,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_xyz.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    },
    "keypoints_uv": {
      "shape": [
        512,
        2,
        21,
        2
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_uv.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    }
  },
  "media_valid": {
    "rgb": true,
    "depth_true_pixels": 102400,
    "depth_pixels": 102400,
    "depth_all_true": true
  },
  "projection": {
    "annotated": true,
    "valid_points": 42,
    "behind_plane_points": 0,
    "near_plane_points": 0,
    "in_frame_points": 21,
    "outside_frame_points": 21,
    "visible_outside_frame": 0,
    "uv_compared": 21,
    "uv_median_px": 0.0321,
    "uv_max_px": 0.0553,
    "uv_within_1px": 21,
    "uv_within_tolerance": 21,
    "outside_uv_max_px": 0.0278,
    "outside_uv_max_relative": 0.000424
  }
}

帧 145 / 512

原生 RGB（MKV 解出的第 145 帧，未重编码）
分割区域（左手红 749 px / 右手蓝 1948 px，叠加在 RGB 上；区域号 2–33，背景号 65534）
原生 UV 关键点（发布的 2D 像素坐标；实心＝源标为可见（基本等于落在画内）、空心＝源标为不可见（画外或被遮挡）；左手 21 点、右手 21 点在画内）
XYZ 投影（相机系米制经本视图自己的 K；左手 21 点、右手 21 点在画内，实心＝源标为可见）；与发布 uv 逐点比对（画内 42 点）：中位偏离 0.0412 px、最大 0.1224 px，≤1 px 的 42 点
z 深度（float32 位串，逐图拉伸；显示范围 0.486–5.000 m）
帧号、来源、全部数值组件与有效性
{
  "frame_index": 145,
  "source": {
    "frame_index": 145,
    "source_frame_id": "1169",
    "timestamp_ns": null,
    "clock_domain": null,
    "source_locator": {
      "archive": "raw/data/RHD_v1-1.zip",
      "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle",
      "source_frame_id": "1169",
      "annotation_index": 1169,
      "member": "RHD_published_v2/evaluation/color/01169.png"
    },
    "sample_id": "RHD/evaluation/01169"
  },
  "components": {
    "intrinsic": {
      "shape": [
        512,
        3,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 512,
      "valid_total": 512,
      "nan_count": 0
    },
    "keypoints_xyz": {
      "shape": [
        512,
        2,
        21,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_xyz.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    },
    "keypoints_uv": {
      "shape": [
        512,
        2,
        21,
        2
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_uv.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    }
  },
  "media_valid": {
    "rgb": true,
    "depth_true_pixels": 102400,
    "depth_pixels": 102400,
    "depth_all_true": true
  },
  "projection": {
    "annotated": true,
    "valid_points": 42,
    "behind_plane_points": 0,
    "near_plane_points": 0,
    "in_frame_points": 42,
    "outside_frame_points": 0,
    "visible_outside_frame": 0,
    "uv_compared": 42,
    "uv_median_px": 0.0412,
    "uv_max_px": 0.1224,
    "uv_within_1px": 42,
    "uv_within_tolerance": 42,
    "outside_uv_max_px": null,
    "outside_uv_max_relative": null
  }
}

帧 256 / 512

原生 RGB（MKV 解出的第 256 帧，未重编码）
分割区域（左手红 2096 px / 右手蓝 0 px，叠加在 RGB 上；区域号 2–33，背景号 65534）
原生 UV 关键点（发布的 2D 像素坐标；实心＝源标为可见（基本等于落在画内）、空心＝源标为不可见（画外或被遮挡）；左手 21 点、右手 0 点在画内）
XYZ 投影（相机系米制经本视图自己的 K；左手 21 点、右手 0 点在画内，实心＝源标为可见）；与发布 uv 逐点比对（画内 21 点）：中位偏离 0.0465 px、最大 0.0679 px，≤1 px 的 21 点；画外 21 点按自身量级的最大相对偏离 0.000264
z 深度（float32 位串，逐图拉伸；显示范围 0.552–5.000 m）
帧号、来源、全部数值组件与有效性
{
  "frame_index": 256,
  "source": {
    "frame_index": 256,
    "source_frame_id": "1280",
    "timestamp_ns": null,
    "clock_domain": null,
    "source_locator": {
      "archive": "raw/data/RHD_v1-1.zip",
      "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle",
      "source_frame_id": "1280",
      "annotation_index": 1280,
      "member": "RHD_published_v2/evaluation/color/01280.png"
    },
    "sample_id": "RHD/evaluation/01280"
  },
  "components": {
    "intrinsic": {
      "shape": [
        512,
        3,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 512,
      "valid_total": 512,
      "nan_count": 0
    },
    "keypoints_xyz": {
      "shape": [
        512,
        2,
        21,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_xyz.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    },
    "keypoints_uv": {
      "shape": [
        512,
        2,
        21,
        2
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_uv.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    }
  },
  "media_valid": {
    "rgb": true,
    "depth_true_pixels": 102400,
    "depth_pixels": 102400,
    "depth_all_true": true
  },
  "projection": {
    "annotated": true,
    "valid_points": 42,
    "behind_plane_points": 0,
    "near_plane_points": 0,
    "in_frame_points": 21,
    "outside_frame_points": 21,
    "visible_outside_frame": 0,
    "uv_compared": 21,
    "uv_median_px": 0.0465,
    "uv_max_px": 0.0679,
    "uv_within_1px": 21,
    "uv_within_tolerance": 21,
    "outside_uv_max_px": 0.0796,
    "outside_uv_max_relative": 0.000264
  }
}

帧 384 / 512

原生 RGB（MKV 解出的第 384 帧，未重编码）
分割区域（左手红 674 px / 右手蓝 3882 px，叠加在 RGB 上；区域号 2–33，背景号 65534）
原生 UV 关键点（发布的 2D 像素坐标；实心＝源标为可见（基本等于落在画内）、空心＝源标为不可见（画外或被遮挡）；左手 3 点、右手 21 点在画内）
XYZ 投影（相机系米制经本视图自己的 K；左手 3 点、右手 21 点在画内，实心＝源标为可见）；与发布 uv 逐点比对（画内 24 点）：中位偏离 0.0345 px、最大 0.0664 px，≤1 px 的 24 点；画外 18 点按自身量级的最大相对偏离 0.00021
z 深度（float32 位串，逐图拉伸；显示范围 0.437–5.000 m）
帧号、来源、全部数值组件与有效性
{
  "frame_index": 384,
  "source": {
    "frame_index": 384,
    "source_frame_id": "1408",
    "timestamp_ns": null,
    "clock_domain": null,
    "source_locator": {
      "archive": "raw/data/RHD_v1-1.zip",
      "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle",
      "source_frame_id": "1408",
      "annotation_index": 1408,
      "member": "RHD_published_v2/evaluation/color/01408.png"
    },
    "sample_id": "RHD/evaluation/01408"
  },
  "components": {
    "intrinsic": {
      "shape": [
        512,
        3,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 512,
      "valid_total": 512,
      "nan_count": 0
    },
    "keypoints_xyz": {
      "shape": [
        512,
        2,
        21,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_xyz.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    },
    "keypoints_uv": {
      "shape": [
        512,
        2,
        21,
        2
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_uv.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    }
  },
  "media_valid": {
    "rgb": true,
    "depth_true_pixels": 102400,
    "depth_pixels": 102400,
    "depth_all_true": true
  },
  "projection": {
    "annotated": true,
    "valid_points": 42,
    "behind_plane_points": 0,
    "near_plane_points": 0,
    "in_frame_points": 24,
    "outside_frame_points": 18,
    "visible_outside_frame": 0,
    "uv_compared": 24,
    "uv_median_px": 0.0345,
    "uv_max_px": 0.0664,
    "uv_within_1px": 24,
    "uv_within_tolerance": 24,
    "outside_uv_max_px": 0.075,
    "outside_uv_max_relative": 0.00021
  }
}

帧 511 / 512

原生 RGB（MKV 解出的第 511 帧，未重编码）
分割区域（左手红 0 px / 右手蓝 2223 px，叠加在 RGB 上；区域号 2–33，背景号 65534）
原生 UV 关键点（发布的 2D 像素坐标；实心＝源标为可见（基本等于落在画内）、空心＝源标为不可见（画外或被遮挡）；左手 0 点、右手 21 点在画内）
XYZ 投影（相机系米制经本视图自己的 K；左手 0 点、右手 21 点在画内，实心＝源标为可见）；与发布 uv 逐点比对（画内 21 点）：中位偏离 0.0267 px、最大 0.0643 px，≤1 px 的 21 点；画外 21 点按自身量级的最大相对偏离 0.000218
z 深度（float32 位串，逐图拉伸；显示范围 0.585–5.000 m）
帧号、来源、全部数值组件与有效性
{
  "frame_index": 511,
  "source": {
    "frame_index": 511,
    "source_frame_id": "1535",
    "timestamp_ns": null,
    "clock_domain": null,
    "source_locator": {
      "archive": "raw/data/RHD_v1-1.zip",
      "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle",
      "source_frame_id": "1535",
      "annotation_index": 1535,
      "member": "RHD_published_v2/evaluation/color/01535.png"
    },
    "sample_id": "RHD/evaluation/01535"
  },
  "components": {
    "intrinsic": {
      "shape": [
        512,
        3,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 512,
      "valid_total": 512,
      "nan_count": 0
    },
    "keypoints_xyz": {
      "shape": [
        512,
        2,
        21,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_xyz.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    },
    "keypoints_uv": {
      "shape": [
        512,
        2,
        21,
        2
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_uv.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    }
  },
  "media_valid": {
    "rgb": true,
    "depth_true_pixels": 102400,
    "depth_pixels": 102400,
    "depth_all_true": true
  },
  "projection": {
    "annotated": true,
    "valid_points": 42,
    "behind_plane_points": 0,
    "near_plane_points": 0,
    "in_frame_points": 21,
    "outside_frame_points": 21,
    "visible_outside_frame": 0,
    "uv_compared": 21,
    "uv_median_px": 0.0267,
    "uv_max_px": 0.0643,
    "uv_within_1px": 21,
    "uv_within_tolerance": 21,
    "outside_uv_max_px": 0.0822,
    "outside_uv_max_relative": 0.000218
  }
}
全部组件元信息及原始来源
{
  "schema_version": "hand-warehouse/1.1",
  "dataset_id": "RHD",
  "release_id": "shared-v1.1",
  "sequence_id": "evaluation_00001024",
  "camera_id": "camera",
  "sample_type": "image",
  "num_frames": 512,
  "hand_slots": [
    "left",
    "right"
  ],
  "sampling": {
    "type": "independent",
    "source_fps": null,
    "frames_ref": "camera/frames.jsonl",
    "selection": "all source independent images sorted by source ID",
    "shard_max_images": 512
  },
  "camera": {
    "model": "pinhole",
    "schema_id": "camera/pinhole/v1",
    "image_size": [
      320,
      320
    ]
  },
  "components": {
    "rgb": {
      "available": true,
      "schema_id": "rgb/uint8/v1",
      "data_ref": "observation/rgb.mkv",
      "axes": [
        "T",
        "height",
        "width",
        "channel"
      ],
      "valid_ref": "observation/rgb_valid.pt",
      "valid_axes": [
        "T"
      ],
      "provenance": {
        "kind": "format_conversion",
        "method": "source values retained"
      },
      "source_locator": {
        "archive": "raw/data/RHD_v1-1.zip",
        "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle"
      },
      "storage": {
        "profile": "mkv/ffv1_rgb/v1",
        "codec": "ffv1",
        "pixel_format": "bgr0",
        "encoded_size": [
          320,
          320
        ],
        "source_lossy": false,
        "decoder": "ffmpeg version 7.1-essentials_build-www.gyan.dev Copyright (c) 2000-2024 the FFmpeg developers",
        "encoder_parameters": [
          "-level",
          "3"
        ]
      },
      "storage_fps": 1,
      "timebase_role": "index_only"
    },
    "intrinsic": {
      "available": true,
      "schema_id": "camera/pinhole/v1",
      "data_ref": "camera/camera.pt",
      "axes": [
        "T",
        "row",
        "column"
      ],
      "valid_ref": "camera/camera.pt",
      "valid_axes": [
        "T"
      ],
      "provenance": {
        "kind": "format_conversion",
        "method": "source values retained"
      },
      "source_locator": {
        "archive": "raw/data/RHD_v1-1.zip",
        "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle"
      },
      "data_key": [
        "intrinsic",
        "data"
      ],
      "valid_key": [
        "intrinsic",
        "valid"
      ],
      "storage": {
        "profile": "pt/tensor_bundle/v1"
      },
      "valid_storage": {
        "profile": "pt/tensor_bundle/v1"
      }
    },
    "keypoints_xyz": {
      "available": true,
      "schema_id": "keypoints/xyz/v1",
      "data_ref": "annotation/keypoints/keypoints.pt",
      "axes": [
        "T",
        "hand",
        "joint",
        "coordinate"
      ],
      "valid_ref": "annotation/keypoints/keypoints.pt",
      "valid_axes": [
        "T",
        "hand",
        "joint"
      ],
      "provenance": {
        "kind": "format_conversion",
        "method": "Keep both supplied hands including invisible points; reorder tip-to-palm to registered proximal-to-tip order"
      },
      "source_locator": {
        "archive": "raw/data/RHD_v1-1.zip",
        "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle"
      },
      "unit": "m",
      "keypoint_schema": "rhd21/v1",
      "coordinate_frame": "camera",
      "source_index_map": [
        0,
        4,
        3,
        2,
        1,
        8,
        7,
        6,
        5,
        12,
        11,
        10,
        9,
        16,
        15,
        14,
        13,
        20,
        19,
        18,
        17
      ],
      "attributes": {
        "visibility": {
          "data_ref": "annotation/keypoints/keypoints.pt",
          "axes": [
            "T",
            "hand",
            "joint"
          ],
          "dtype": "bool",
          "visibility_type": "boolean",
          "meaning": "source per-keypoint visibility",
          "provenance": {
            "kind": "format_conversion",
            "method": "native uv_vis visibility reordered with coordinates"
          },
          "source_locator": {
            "archive": "raw/data/RHD_v1-1.zip",
            "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle"
          },
          "data_key": [
            "visibility",
            "data"
          ],
          "valid_ref": "annotation/keypoints/keypoints.pt",
          "valid_key": [
            "visibility",
            "valid"
          ],
          "storage": {
            "profile": "pt/tensor_bundle/v1"
          },
          "valid_storage": {
            "profile": "pt/tensor_bundle/v1"
          },
          "valid_axes": [
            "T",
            "hand",
            "joint"
          ]
        }
      },
      "data_key": [
        "xyz",
        "data"
      ],
      "valid_key": [
        "xyz",
        "valid"
      ],
      "storage": {
        "profile": "pt/tensor_bundle/v1"
      },
      "valid_storage": {
        "profile": "pt/tensor_bundle/v1"
      }
    },
    "keypoints_uv": {
      "available": true,
      "schema_id": "keypoints/uv/v1",
      "data_ref": "annotation/keypoints/keypoints.pt",
      "axes": [
        "T",
        "hand",
        "joint",
        "coordinate"
      ],
      "valid_ref": "annotation/keypoints/keypoints.pt",
      "valid_axes": [
        "T",
        "hand",
        "joint"
      ],
      "provenance": {
        "kind": "format_conversion",
        "method": "Keep both supplied hands including invisible points; reorder tip-to-palm to registered proximal-to-tip order"
      },
      "source_locator": {
        "archive": "raw/data/RHD_v1-1.zip",
        "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle"
      },
      "unit": "pixel",
      "keypoint_schema": "rhd21/v1",
      "coordinate_frame": "rgb_image",
      "source_index_map": [
        0,
        4,
        3,
        2,
        1,
        8,
        7,
        6,
        5,
        12,
        11,
        10,
        9,
        16,
        15,
        14,
        13,
        20,
        19,
        18,
        17
      ],
      "attributes": {
        "visibility": {
          "data_ref": "annotation/keypoints/keypoints.pt",
          "axes": [
            "T",
            "hand",
            "joint"
          ],
          "dtype": "bool",
          "visibility_type": "boolean",
          "meaning": "source per-keypoint visibility",
          "provenance": {
            "kind": "format_conversion",
            "method": "native uv_vis visibility reordered with coordinates"
          },
          "source_locator": {
            "archive": "raw/data/RHD_v1-1.zip",
            "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle"
          },
          "data_key": [
            "visibility",
            "data"
          ],
          "valid_ref": "annotation/keypoints/keypoints.pt",
          "valid_key": [
            "visibility",
            "valid"
          ],
          "storage": {
            "profile": "pt/tensor_bundle/v1"
          },
          "valid_storage": {
            "profile": "pt/tensor_bundle/v1"
          },
          "valid_axes": [
            "T",
            "hand",
            "joint"
          ]
        }
      },
      "data_key": [
        "uv",
        "data"
      ],
      "valid_key": [
        "uv",
        "valid"
      ],
      "storage": {
        "profile": "pt/tensor_bundle/v1"
      },
      "valid_storage": {
        "profile": "pt/tensor_bundle/v1"
      }
    },
    "depth": {
      "available": true,
      "schema_id": "depth/z_float32/v1",
      "data_ref": "observation/depth.mkv",
      "axes": [
        "T",
        "height",
        "width"
      ],
      "valid_ref": "observation/depth_valid.mkv",
      "valid_axes": [
        "T",
        "height",
        "width"
      ],
      "storage": {
        "profile": "mkv/depth_float32_bits/v1",
        "codec": "ffv1",
        "pixel_format": "gray16le",
        "encoded_size": [
          640,
          320
        ],
        "source_lossy": false,
        "decoder": "ffmpeg version 7.1-essentials_build-www.gyan.dev Copyright (c) 2000-2024 the FFmpeg developers",
        "encoder_parameters": [
          "-level",
          "3"
        ]
      },
      "valid_storage": {
        "profile": "mkv/bool_ffv1/v1",
        "codec": "ffv1",
        "pixel_format": "gray",
        "encoded_size": [
          320,
          320
        ],
        "source_lossy": false,
        "decoder": "ffmpeg version 7.1-essentials_build-www.gyan.dev Copyright (c) 2000-2024 the FFmpeg developers",
        "encoder_parameters": [
          "-level",
          "3"
        ]
      },
      "unit": "m",
      "depth_type": "z",
      "depth_origin": "rendered",
      "coordinate_frame": "current_rgb_camera",
      "pixel_grid": "current_rgb",
      "storage_fps": 1,
      "timebase_role": "index_only",
      "provenance": {
        "kind": "format_conversion",
        "method": "Official code (R*256+G)/65535*5 m; preserve complete float32 bits; source supplies every pixel, no additional invalid-pixel convention inferred"
      },
      "source_locator": {
        "archive": "raw/data/RHD_v1-1.zip",
        "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle",
        "member_pattern": "RHD_published_v2/evaluation/depth/{source_frame_id:05d}.png"
      }
    },
    "segment": {
      "available": true,
      "schema_id": "segment/uint16_labels/v1",
      "data_ref": "annotation/segment/mask.mkv",
      "axes": [
        "T",
        "height",
        "width"
      ],
      "valid_ref": "annotation/segment/mask_valid.pt",
      "valid_axes": [
        "T"
      ],
      "provenance": {
        "kind": "format_conversion",
        "method": "Native codes preserved; source 0 background mapped to 65534; native fine-region codes retained without merging"
      },
      "source_locator": {
        "archive": "raw/data/RHD_v1-1.zip",
        "annotation": "RHD_published_v2/evaluation/anno_evaluation.pickle"
      },
      "storage": {
        "profile": "mkv/uint16_ffv1/v1",
        "codec": "ffv1",
        "pixel_format": "gray16le",
        "encoded_size": [
          320,
          320
        ],
        "source_lossy": false,
        "decoder": "ffmpeg version 7.1-essentials_build-www.gyan.dev Copyright (c) 2000-2024 the FFmpeg developers",
        "encoder_parameters": [
          "-level",
          "3"
        ]
      },
      "storage_fps": 1,
      "timebase_role": "index_only",
      "segmentation_type": "semantic",
      "mapping_ref": "annotation/segment/instances.jsonl",
      "coverage": {
        "kind": "full",
        "scope": "official rendered background, person and both hand regions"
      }
    }
  },
  "split": "test",
  "storage_revision": "lossless-video-2026-09-18",
  "split_scheme": "RHD_official_training_evaluation",
  "split_source_locator": {
    "archive": "raw/data/RHD_v1-1.zip",
    "member_prefix": "RHD_published_v2/evaluation"
  },
  "view_type": "exo"
}
持久化 QC 结果
{
  "view": "evaluation_00001024/camera",
  "sequence_id": "evaluation_00001024",
  "camera_id": "camera",
  "view_type": "exo",
  "split": "test",
  "num_frames": 512,
  "camera": {
    "model": "pinhole",
    "schema_id": "camera/pinhole/v1",
    "image_size": [
      320,
      320
    ]
  },
  "numeric_components": {
    "intrinsic": {
      "shape": [
        512,
        3,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 512,
      "valid_total": 512,
      "nan_count": 0
    },
    "keypoints_xyz": {
      "shape": [
        512,
        2,
        21,
        3
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_xyz.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    },
    "keypoints_uv": {
      "shape": [
        512,
        2,
        21,
        2
      ],
      "dtype": "torch.float32",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": 0
    },
    "keypoints_uv.visibility": {
      "shape": [
        512,
        2,
        21
      ],
      "dtype": "torch.bool",
      "valid_true": 21504,
      "valid_total": 21504,
      "nan_count": null
    }
  },
  "bundle_files": 2,
  "bundles": [
    "annotation/keypoints/keypoints.pt",
    "camera/camera.pt"
  ],
  "media": [
    {
      "component": "rgb",
      "ref": "observation/rgb.mkv",
      "frames_decoded": 6
    },
    {
      "component": "depth",
      "ref": "observation/depth.mkv",
      "frames_decoded": 6
    },
    {
      "component": "segment",
      "ref": "annotation/segment/mask.mkv",
      "frames_decoded": 6
    }
  ],
  "consumed_files": [
    "annotation/keypoints/keypoints.pt",
    "annotation/segment/mask.mkv",
    "camera/camera.pt",
    "observation/depth.mkv",
    "observation/rgb.mkv"
  ],
  "selected_frames": [
    0,
    128,
    145,
    256,
    384,
    511
  ],
  "decode_scope": "ffmpeg select= demux filter over the preserved MKV; only the frames listed in selected_frames are decoded, each checked against its expected byte count",
  "projection": {
    "model": "pinhole",
    "calibrated": true,
    "coordinate_frame": "camera",
    "scope": "saved camera-frame xyz (metres) through the view's own pinhole intrinsic, compared against the saved native uv; not a media pixel comparison",
    "uv_tolerance_px": 1.0,
    "visibility_rule": "发布的 uv 是原始投影值，画外关节的坐标可以任意大；本页按发布的 uv 是否落在画内分两组核验——画内点按像素偏差、画外点按相对自身量级的偏差，另单列\"源标为可见却落在画外\"的点数（该方向才可能是标注与标定不一致）",
    "view_scan": {
      "frames_scanned": 512,
      "points_scanned": 21504,
      "behind_plane_points": 331,
      "near_plane_points": 4,
      "comparable_points": 21169,
      "in_frame_points": 15642,
      "outside_frame_points": 5527,
      "median_px": 0.0382,
      "in_frame_max_px": 0.1224,
      "frames_with_in_frame_points": 512,
      "frames_all_in_frame_points_within_1px": 512,
      "outside_frame_within_1px": 5458,
      "outside_frame_max_px": 41.5559,
      "outside_frame_max_px_uv_scale": 151400.0,
      "outside_frame_max_relative": 0.0031,
      "outside_frame_max_relative_px": 0.0741,
      "outside_frame_max_relative_uv_scale": 23.9,
      "worst_frame": 145,
      "worst_frame_in_frame_max_px": 0.1224,
      "visibility_true": 15181,
      "visible_outside_frame": 0,
      "visible_outside_frame_max_px": null,
      "visible_near_plane_points": 0,
      "visible_behind_plane_points": 0,
      "invisible_in_frame": 461,
      "plane_threshold_m": 0.001
    },
    "points": 252,
    "inside": 159,
    "off_image": 93,
    "uv_compared": 159,
    "uv_within_1px": 159,
    "uv_within_tolerance": 159,
    "frames": {
      "0": {
        "annotated": true,
        "valid_points": 42,
        "behind_plane_points": 0,
        "near_plane_points": 0,
        "in_frame_points": 30,
        "outside_frame_points": 12,
        "visible_outside_frame": 0,
        "uv_compared": 30,
        "uv_median_px": 0.0303,
        "uv_max_px": 0.0614,
        "uv_within_1px": 30,
        "uv_within_tolerance": 30,
        "outside_uv_max_px": 0.028,
        "outside_uv_max_relative": 0.000341
      },
      "128": {
        "annotated": true,
        "valid_points": 42,
        "behind_plane_points": 0,
        "near_plane_points": 0,
        "in_frame_points": 21,
        "outside_frame_points": 21,
        "visible_outside_frame": 0,
        "uv_compared": 21,
        "uv_median_px": 0.0321,
        "uv_max_px": 0.0553,
        "uv_within_1px": 21,
        "uv_within_tolerance": 21,
        "outside_uv_max_px": 0.0278,
        "outside_uv_max_relative": 0.000424
      },
      "145": {
        "annotated": true,
        "valid_points": 42,
        "behind_plane_points": 0,
        "near_plane_points": 0,
        "in_frame_points": 42,
        "outside_frame_points": 0,
        "visible_outside_frame": 0,
        "uv_compared": 42,
        "uv_median_px": 0.0412,
        "uv_max_px": 0.1224,
        "uv_within_1px": 42,
        "uv_within_tolerance": 42,
        "outside_uv_max_px": null,
        "outside_uv_max_relative": null
      },
      "256": {
        "annotated": true,
        "valid_points": 42,
        "behind_plane_points": 0,
        "near_plane_points": 0,
        "in_frame_points": 21,
        "outside_frame_points": 21,
        "visible_outside_frame": 0,
        "uv_compared": 21,
        "uv_median_px": 0.0465,
        "uv_max_px": 0.0679,
        "uv_within_1px": 21,
        "uv_within_tolerance": 21,
        "outside_uv_max_px": 0.0796,
        "outside_uv_max_relative": 0.000264
      },
      "384": {
        "annotated": true,
        "valid_points": 42,
        "behind_plane_points": 0,
        "near_plane_points": 0,
        "in_frame_points": 24,
        "outside_frame_points": 18,
        "visible_outside_frame": 0,
        "uv_compared": 24,
        "uv_median_px": 0.0345,
        "uv_max_px": 0.0664,
        "uv_within_1px": 24,
        "uv_within_tolerance": 24,
        "outside_uv_max_px": 0.075,
        "outside_uv_max_relative": 0.00021
      },
      "511": {
        "annotated": true,
        "valid_points": 42,
        "behind_plane_points": 0,
        "near_plane_points": 0,
        "in_frame_points": 21,
        "outside_frame_points": 21,
        "visible_outside_frame": 0,
        "uv_compared": 21,
        "uv_median_px": 0.0267,
        "uv_max_px": 0.0643,
        "uv_within_1px": 21,
        "uv_within_tolerance": 21,
        "outside_uv_max_px": 0.0822,
        "outside_uv_max_relative": 0.000218
      }
    },
    "note": "RHD publishes uv and xyz with one pinhole K, so this is the published projection_qc rather than a one-sided count: the saved xyz is projected through the view's own intrinsic and compared with the saved uv, split into the joints whose published uv lies inside the picture (pixel tolerance) and those outside it (relative to their own magnitude). view_scan covers every frame of the view; the per-frame entries below cover only the sampled frames whose pixels were decoded."
  },
  "scope": "v1.1 registry, keys, shapes, validity and timeline; not a media pixel comparison",
  "not_drawn": [],
  "visible_outside_frame_frames": [],
  "not_annotated_frames": []
}