| |
| """生成 data/datasets.json。改数据集信息就改这个文件，然后重新运行： |
| python tools/make_data.py |
| """ |
| import json |
| from pathlib import Path |
|
|
| ROOT = Path(__file__).resolve().parent.parent |
|
|
| TEAM = "https://yangzijing-nine-source-hand-data-browser.static.hf.space" |
| T_WILOR = {"label": "团队图册：WiLoR 14 源 + 合成数据", "url": f"{TEAM}/hand_dataset_browser_20261013/index.html"} |
| T_EGO = {"label": "团队图册：第一人称具身数据集", "url": f"{TEAM}/egocentric_embodied_datasets_20261020/index.html"} |
| T_2D = {"label": "团队图册：2D 数据集", "url": f"{TEAM}/hand_2d_dataset_browser_20261013/index.html"} |
| T_VID = {"label": "团队视频：九源 MANO GT 叠加（20s）", "url": "https://huggingface.co/datasets/yangzijing/nine-source-hand-data-review-videos-20s/tree/main"} |
|
|
|
|
| def arxiv(i, label="论文"): |
| return {"label": f"{label} arXiv:{i}", "url": f"https://arxiv.org/abs/{i}"} |
|
|
|
|
| meta = { |
| "title": "SOMA 手部数据图谱", |
| "updated": "2026-11-02", |
| "goal": "从单目 RGB 恢复 SOMA hand 的 shape 与 pose：按数据来源和标注方法整理可用的手部数据集。", |
| "sources": [ |
| {"id": "ego", "name": "第一视角", "desc": "头戴或胸前相机拍的真实视频"}, |
| {"id": "egoexo", "name": "第一视角 + 外视角", "desc": "同步的头戴相机和固定相机"}, |
| {"id": "exo", "name": "外视角多相机", "desc": "棚内固定多相机，没有头戴视角"}, |
| {"id": "scan", "name": "扫描仪", "desc": "静态手势的多相机三维扫描"}, |
| {"id": "wild", "name": "野外图像 / 视频", "desc": "网络或日常拍摄，没有标定装置"}, |
| {"id": "synth", "name": "合成渲染", "desc": "渲染或重光照生成的图像"}, |
| {"id": "norgb", "name": "无 RGB", "desc": "只有动捕、事件相机或黑白相机"}, |
| ], |
| "methods": [ |
| {"id": "scan", "name": "三维扫描配准", "short": "扫描", |
| "steps": ["多相机扫描仪拍静态手势", "重建亚毫米级网格", "关键点三角化", "配准 MANO"], |
| "trust": "最高。直接测得手的表面，是目前最可靠的 shape 真值。", |
| "issues": "只有静态手势，没有手物交互，也没有第一视角。", |
| "route": "R6"}, |
| {"id": "marker", "name": "光学 marker 动捕 + 拟合", "short": "光学动捕", |
| "steps": ["手上贴反光 marker", "Vicon / OptiTrack 跟踪", "逐人标定 shape", "拟合 MANO 或 SMPL-X"], |
| "trust": "高。shape 逐人标定，pose 不受相机遮挡影响。", |
| "issues": "marker 会出现在 RGB 里；人数通常很少；重遮挡时 marker 丢失。", |
| "route": "R1"}, |
| {"id": "rgbd", "name": "多视角 RGB-D 拟合", "short": "RGB-D 拟合", |
| "steps": ["多台 RGB-D 同步拍摄", "2D 关键点（人工或模型）+ 深度", "多视角联合优化", "拟合 MANO，人工核验"], |
| "trust": "中。受深度噪声和 2D 精度限制，shape 可用但偏粗。", |
| "issues": "物体遮挡时手指误差大；HOnnotate 一类方法依赖初始化。", |
| "route": "R1"}, |
| {"id": "mvkp", "name": "多视角关键点三角化", "short": "关键点三角化", |
| "steps": ["多相机 RGB", "2D 关键点（人工或模型）", "三角化得到 3D 关键点", "可选：拟合 MANO"], |
| "trust": "中低。shape 只由关键点约束；只给关键点时没有 shape。", |
| "issues": "模型预标注会带入 HaMeR 等方法的偏差。", |
| "route": "R1 / R3"}, |
| {"id": "device", "name": "设备追踪", "short": "设备追踪", |
| "steps": ["头显、手套或 IMU", "设备内置手部追踪", "输出关节或设备自带手模型"], |
| "trust": "低。精度依赖设备，通常没有公开误差。", |
| "issues": "关节定义因设备而异，需要逐个对齐。", |
| "route": "R3"}, |
| {"id": "synth", "name": "合成渲染", "short": "合成", |
| "steps": ["参数化手 + 动作库", "渲染（光照、材质、相机）", "直接导出真值"], |
| "trust": "真值精确，但 shape 分布取决于采样方式。", |
| "issues": "外观和光照与真实图像有域差。", |
| "route": "R1 / R2"}, |
| {"id": "manual2d", "name": "人工 2D 标注", "short": "人工 2D", |
| "steps": ["图像", "模型预标注（可选）", "人工标注或修正 2D 关键点"], |
| "trust": "没有 3D 信息，只提供 2D 监督。", |
| "issues": "遮挡点的标注口径不一。", |
| "route": "R4"}, |
| {"id": "pseudo", "name": "单目模型伪标签", "short": "伪标签", |
| "steps": ["单目视频", "HaMeR 或自研管线预测 MANO", "可选：修复与筛选"], |
| "trust": "最低。不能当 shape 真值，只适合弱监督或预训练。", |
| "issues": "继承预测模型自身的系统偏差。", |
| "route": "R7"}, |
| ], |
| "tiers": [ |
| {"id": "S1", "name": "扫描配准", "desc": "直接测得表面"}, |
| {"id": "S2", "name": "动捕逐人标定", "desc": "marker 约束的逐人 shape"}, |
| {"id": "S3", "name": "RGB-D 拟合", "desc": "深度约束，偏粗"}, |
| {"id": "S4", "name": "关键点拟合", "desc": "只靠关键点约束"}, |
| {"id": "S5", "name": "伪标签", "desc": "不可作为 shape 真值"}, |
| {"id": "SY", "name": "合成精确", "desc": "精确，但分布看采样"}, |
| {"id": "-", "name": "无 shape", "desc": "只有关节或 2D"}, |
| ], |
| "statuses": [ |
| {"id": "used", "name": "已在使用", "desc": "九源、WiLoR 14 源或基准中已经用到"}, |
| {"id": "surveyed", "name": "团队已调研", "desc": "已在团队图册中，尚未进训练"}, |
| {"id": "retired", "name": "不再使用", "desc": "当前不再使用"}, |
| {"id": "p0", "name": "P0 新增", "desc": "建议优先接入"}, |
| {"id": "p1", "name": "P1 新增", "desc": "P0 之后接入"}, |
| {"id": "p2", "name": "P2 新增", "desc": "有余力时接入"}, |
| {"id": "own", "name": "自采", "desc": "我们自己的采集"}, |
| {"id": "no", "name": "不建议", "desc": "不满足真实 RGB 或真值条件"}, |
| ], |
| "routes": [ |
| {"id": "R0", "name": "原生 SOMA", "from": "SOMA 拟合结果", "how": "直接使用。", "shape": True}, |
| {"id": "R1", "name": "MANO → SOMA", "from": "MANO 参数或网格", |
| "how": "用 SOMAHandLayer 的 MANO 后端映射到 SOMA 拓扑，再重拟合 SOMA 原生 20 维 identity 和 pose；按数据集统计顶点残差，残差大的帧不监督 shape。", |
| "shape": True}, |
| {"id": "R2", "name": "SMPL-X → SOMA", "from": "SMPL-X / SMPL-H", |
| "how": "取手部 MANO 部分走 R1；或用 SOMA-X 的 SMPL-X 身体后端整体转换。手型跟随身体 shape，分布偏窄。", |
| "shape": True}, |
| {"id": "R3", "name": "3D 关键点 → SOMA pose", "from": "21 / 25 点 3D 关节", |
| "how": "先定 MANO 21 点（指尖为网格顶点）与 SOMA 关节的对应，只拟合 pose；shape 固定，不参与监督。", |
| "shape": False}, |
| {"id": "R4", "name": "2D 关键点", "from": "2D 关键点", |
| "how": "不转换。训练时用重投影损失监督。", |
| "shape": False}, |
| {"id": "R5", "name": "MHR → SOMA", "from": "MHR 参数", |
| "how": "SOMAHandLayer 原生支持 MHR 后端，可直接映射。SAM 3D Body 输出 MHR，可当野外图像的伪标签教师。", |
| "shape": True}, |
| {"id": "R6", "name": "扫描 → SOMA", "from": "扫描网格", |
| "how": "直接把 SOMA 原生 shape + pose 拟合到扫描网格，是最好的 shape 监督和评测来源。", |
| "shape": True}, |
| {"id": "R7", "name": "伪标签", "from": "预测的 MANO", |
| "how": "可走 R1 转换，但只作弱监督或预训练，不监督 shape。", |
| "shape": False}, |
| ], |
| "teamLinks": [T_WILOR, T_EGO, T_2D, T_VID], |
| } |
|
|
| D = [] |
|
|
|
|
| def add(**kw): |
| kw.setdefault("tags", []) |
| kw.setdefault("links", []) |
| kw.setdefault("team", []) |
| kw.setdefault("risk", "") |
| kw.setdefault("subjects", "") |
| kw.setdefault("scale", "") |
| kw.setdefault("year", "") |
| kw.setdefault("rig", {}) |
| D.append(kw) |
|
|
|
|
| |
| add(id="arctic", name="ARCTIC", year="CVPR 2023", status="used", tags=["九源", "WiLoR 14 源", "基准"], |
| source="egoexo", method="marker", gt="MANO + SMPL-X", tier="S2", route="R1", |
| camera="8 个固定 RGB + 1 个第一视角 RGB", rig={"ego": True, "exo": 8}, |
| subjects="10 人", scale="约 210 万张图，339 个序列", |
| fills="双手操作铰接物体，带第一视角和光学真值。", |
| risk="人数少，shape 分布窄。", |
| links=[arxiv("2204.13662")], team=[T_WILOR, T_VID]) |
| add(id="hot3d", name="HOT3D", year="CVPR 2025", status="used", tags=["九源", "WiLoR 14 源", "基准"], |
| source="ego", method="marker", gt="MANO / UmeTrack", tier="S2", route="R1", |
| camera="Aria RGB 鱼眼 + Quest 3 黑白", rig={"ego": True, "fisheye": True}, |
| subjects="19 人", scale="370 万张以上，约 833 分钟", |
| fills="真实第一视角鱼眼数据，光学动捕真值。", |
| risk="我们的模型在这里 RR 偏差大，先核对裁剪和去畸变与相机模型是否一致。", |
| links=[{"label": "项目主页", "url": "https://facebookresearch.github.io/hot3d/"}], team=[T_WILOR, T_VID]) |
| add(id="h2o", name="H2O", year="ICCV 2021", status="used", tags=["九源", "基准"], |
| source="egoexo", method="rgbd", gt="MANO", tier="S3", route="R1", |
| camera="5 台 Azure Kinect RGB-D，其中 1 台头戴", rig={"ego": True, "exo": 4, "depth": True}, |
| subjects="4 人", scale="571,645 张图，184 个序列", |
| fills="头戴相机与 Femto Bolt 同一传感器族，相机域最接近我们的采集。", |
| risk="人数很少。", |
| links=[arxiv("2104.11181")], team=[T_VID]) |
| add(id="hoi4d", name="HOI4D", year="CVPR 2022", status="retired", tags=["九源"], |
| source="ego", method="rgbd", gt="MANO", tier="S3", route="R1", |
| camera="头戴 RGB-D", rig={"ego": True, "depth": True}, |
| subjects="9 人", scale="约 240 万帧，4,000 个序列", |
| fills="第一视角的类别级手物交互。", |
| risk="右手偏置明显，左手只占约 10%。", |
| links=[arxiv("2203.01577")], team=[T_VID]) |
| add(id="dexycb", name="DexYCB", year="CVPR 2021", status="used", tags=["九源", "WiLoR 14 源"], |
| source="exo", method="rgbd", gt="MANO", tier="S3", route="R1", |
| camera="8 台 RGB-D", rig={"exo": 8, "depth": True}, |
| subjects="10 人", scale="约 58.2 万张图，1,000 个序列", |
| fills="抓取 YCB 物体，人工多视角 2D 加深度拟合。", |
| links=[arxiv("2104.04631")], team=[T_WILOR, T_VID]) |
| add(id="ho3d", name="HO3D v3", year="CVPR 2020", status="used", tags=["九源", "WiLoR 14 源"], |
| source="exo", method="rgbd", gt="MANO", tier="S3", route="R1", |
| camera="1 到 5 台 RGB-D", rig={"exo": 5, "depth": True}, |
| subjects="10 人", scale="103,462 张图", |
| fills="单手与物体交互，HOnnotate 联合拟合。", |
| links=[arxiv("1907.01481")], team=[T_WILOR, T_VID]) |
| add(id="h2o3d", name="H2O3D", year="2022", status="used", tags=["WiLoR 14 源"], |
| source="exo", method="rgbd", gt="MANO", tier="S3", route="R1", |
| camera="5 台 RGB-D", rig={"exo": 5, "depth": True}, |
| subjects="待核", scale="76,340 张图，85 个序列", |
| fills="双手与物体交互。", |
| team=[T_WILOR]) |
| add(id="interhand", name="InterHand2.6M", year="ECCV 2020", status="used", tags=["WiLoR 14 源"], |
| source="exo", method="mvkp", gt="MANO", tier="S4", route="R1", |
| camera="棚内多相机 RGB", rig={"exo": 12}, |
| subjects="待核", scale="5 fps 版 2,590,347 张", |
| fills="双手交互和丰富的手指姿态。", |
| links=[arxiv("2008.09309")], team=[T_WILOR]) |
| add(id="reinterhand", name="Re:InterHand", year="NeurIPS 2023", status="used", tags=["WiLoR 14 源"], |
| source="synth", method="mvkp", gt="MANO", tier="S4", route="R1", |
| camera="真实动作重光照渲染，含第一视角", rig={"synth": True, "ego": True}, |
| scale="1,526,795 张图，有效 147,745 帧", |
| fills="真实双手动作配多样光照，含第一视角渲染。", |
| team=[T_WILOR]) |
| add(id="freihand", name="FreiHAND", year="ICCV 2019", status="used", tags=["WiLoR 14 源"], |
| source="exo", method="mvkp", gt="MANO", tier="S4", route="R1", |
| camera="棚内多相机，绿幕换背景", rig={"exo": 8}, |
| subjects="32 人", scale="训练 130,240 张，测试 3,960 张", |
| fills="单手姿态与 shape 的经典基准。", |
| links=[arxiv("1909.04349")], team=[T_WILOR]) |
| add(id="mtc", name="MTC", year="CVPR 2019", status="used", tags=["WiLoR 14 源"], |
| source="exo", method="mvkp", gt="3D 关键点", tier="-", route="R3", |
| camera="Panoptic 棚内多相机", rig={"exo": 12}, |
| subjects="40 人", scale="清洗后约 11.1 万个手样本", |
| fills="全身动作中的手，视角多样。", |
| links=[arxiv("1812.01598")], team=[T_WILOR]) |
| add(id="bedlam", name="BEDLAM", year="CVPR 2023", status="used", tags=["WiLoR 14 源"], |
| source="synth", method="synth", gt="SMPL-X", tier="SY", route="R2", |
| camera="合成渲染，非第一视角", rig={"synth": True}, |
| scale="10,450 个序列，约 160 万张，约 15 小时", |
| fills="全身场景中的手，背景与光照多样。", |
| risk="手型跟随 SMPL-X 身体 shape，分布偏窄。", |
| links=[arxiv("2306.16940")], team=[T_WILOR]) |
| add(id="rhd", name="RHD", year="ICCV 2017", status="used", tags=["WiLoR 14 源"], |
| source="synth", method="synth", gt="3D 关键点", tier="-", route="R3", |
| camera="合成渲染", rig={"synth": True}, |
| scale="训练 41,258 张，测试 2,728 张", |
| fills="早期合成手部关键点数据。", |
| links=[arxiv("1705.01389")], team=[T_WILOR]) |
| add(id="egorenderer", name="EgoRenderer V005 / V006", year="2026", status="used", tags=["自有合成"], |
| source="synth", method="synth", gt="原生手部参数；拟合 MANO / SOMA 为非 GT 辅助标注", tier="SY", route="R1", |
| camera="第一视角合成渲染", rig={"synth": True, "ego": True}, |
| scale="V005：1,314 源组 / 2,425 历史包；V006 ARCTIC：301 源动作，每条只生产 1 份", |
| fills="可控的第一视角合成数据。", |
| risk="V005 参数附件更新沿用原 RGB；V006 渲染使用原生手，不使用拟合参数生成像素。当前只保留每动作一份生产目标，历史额外变体不计入目标。辅助拟合与训练资格另行登记。", |
| team=[T_WILOR]) |
| add(id="cocowb", name="COCO-WholeBody", year="ECCV 2020", status="used", tags=["WiLoR 14 源"], |
| source="wild", method="manual2d", gt="2D 关键点", tier="-", route="R4", |
| camera="网络图像", rig={"wild": True}, |
| scale="约 20 万张图", |
| fills="野外图像的 2D 监督。", |
| links=[arxiv("2007.11858")], team=[T_WILOR]) |
| add(id="halpe", name="Halpe", year="2020", status="used", tags=["WiLoR 14 源"], |
| source="wild", method="manual2d", gt="2D 关键点", tier="-", route="R4", |
| camera="网络图像", rig={"wild": True}, |
| scale="约 4 万张训练图", |
| fills="野外图像的 2D 监督。", |
| team=[T_WILOR]) |
| add(id="mpiinzsl", name="MPII + NZSL", year="CVPR 2017", status="used", tags=["WiLoR 14 源"], |
| source="wild", method="manual2d", gt="2D 关键点", tier="-", route="R4", |
| camera="网络图像与手语视频", rig={"wild": True}, |
| scale="2,800 只人工标注的手", |
| fills="人工 2D 手部关键点。", |
| links=[arxiv("1704.07809")], team=[T_WILOR]) |
| add(id="ego4d", name="Ego4D（P3R3 伪标签）", year="CVPR 2022", status="used", tags=["九源（伪标签）"], |
| source="wild", method="pseudo", gt="预测 MANO", tier="S5", route="R7", |
| camera="多种头戴相机", rig={"ego": True, "wild": True}, |
| fills="海量真实第一视角外观。", |
| risk="P3R3 管线预测的 MANO，不是训练真值。", |
| team=[T_VID]) |
| add(id="epic", name="EPIC-KITCHENS（P3R3 伪标签）", year="2018 / 2022", status="used", tags=["九源（伪标签）"], |
| source="wild", method="pseudo", gt="预测 MANO", tier="S5", route="R7", |
| camera="头戴 GoPro", rig={"ego": True, "wild": True}, |
| fills="厨房场景的真实第一视角。", |
| risk="P3R3 管线预测的 MANO，不是训练真值。", |
| team=[T_VID]) |
| add(id="egodex", name="EgoDex", year="2025", status="used", tags=["九源（伪标签）", "第一人称图册"], |
| source="ego", method="device", gt="25 个 3D 关节（ARKit）", tier="-", route="R3", |
| camera="Apple Vision Pro", rig={"ego": True}, |
| scale="约 829 小时", |
| fills="大规模第一视角灵巧操作，带设备追踪的关节。", |
| risk="关节来自设备追踪，团队另有 P3R3 预测 MANO，都不能监督 shape。", |
| links=[arxiv("2505.11709")], team=[T_EGO, T_VID]) |
|
|
| |
| add(id="egopressure", name="EgoPressure", year="CVPR 2025", status="surveyed", tags=["第一人称图册"], |
| source="egoexo", method="rgbd", gt="MANO", tier="S3", route="R1", |
| camera="头戴相机 + 多台外视角 RGB-D", rig={"ego": True, "exo": 7, "depth": True}, |
| fills="手与表面接触的压力，带第一视角。", |
| risk="HaMeR 初始化后用多视角 RGB-D、分割、压力和时序约束拟合 MANO，手上没有 marker。", |
| links=[arxiv("2409.02224")], team=[T_EGO]) |
| add(id="egoemg", name="EgoEMG", year="2026", status="surveyed", tags=["第一人称图册"], |
| source="ego", method="marker", gt="MANO（光捕 marker 经 markers2mano 拟合）", tier="S2", route="R1", |
| camera="头戴 GoPro 60 FPS + 26 相机光学动捕", rig={"ego": True}, |
| fills="第一视角与肌电同步，光学动捕真值。", |
| links=[arxiv("2605.05712")], team=[T_EGO]) |
| add(id="hocap", name="HO-Cap", year="2024", status="surveyed", tags=["第一人称图册"], |
| source="egoexo", method="rgbd", gt="MANO", tier="S3", route="R1", |
| camera="多台 RGB-D + HoloLens 第一视角 RGB", rig={"ego": True, "exo": 8, "depth": True}, |
| fills="手物交互，含第一视角。", |
| risk="多视角 RGB-D 与关键点约束拟合 MANO，不使用光捕。", |
| team=[T_EGO]) |
| add(id="taco", name="TACO", year="CVPR 2024", status="surveyed", tags=["第一人称图册"], |
| source="egoexo", method="mvkp", gt="MANO", tier="S4", route="R1", |
| camera="多台外视角 + 1 个第一视角", rig={"ego": True, "exo": 12}, |
| fills="双手使用工具的交互。", |
| risk="多视角 MMPose 三角化后做 MANO 接触与时序拟合，手上没有光捕。", |
| team=[T_EGO]) |
| add(id="aceda0", name="ACE-Data-0", year="2026", status="surveyed", tags=["第一人称图册"], |
| source="egoexo", method="marker", gt="手部动捕（格式待核）", tier="S2", route="R1", |
| camera="环境多相机 + 第一视角", rig={"ego": True, "exo": 6}, |
| fills="身体与物体用 OptiTrack，手部附带动捕数据。", |
| risk="官方叠加样例未逐帧配对；全量发布目标 150 小时，目前为快照。", |
| team=[T_EGO]) |
| add(id="xperience", name="Xperience-10M", year="2026", status="surveyed", tags=["第一人称图册"], |
| source="ego", method="device", gt="手部动捕（类型待核）", tier="-", route="R3", |
| camera="第一视角", rig={"ego": True}, |
| scale="约 1 万小时", |
| fills="规模极大的第一视角数据。", |
| risk="手部真值的采集方式和精度待核。", |
| team=[T_EGO]) |
| add(id="feel", name="FEEL", year="2026", status="surveyed", tags=["第一人称图册"], |
| source="ego", method="device", gt="Aria MPS 手部追踪", tier="-", route="R3", |
| camera="Aria", rig={"ego": True, "fisheye": True}, |
| fills="Aria 第一视角的设备追踪手部。", |
| team=[T_EGO]) |
| add(id="openaoe", name="Open-AoE", year="2026", status="surveyed", tags=["第一人称图册"], |
| source="ego", method="pseudo", gt="自动重建 MANO", tier="S5", route="R7", |
| camera="第一视角", rig={"ego": True}, |
| fills="第一视角操作，自动 MANO 重建。", |
| team=[T_EGO]) |
| add(id="genhumanego", name="Gen-HumanEgo", year="2026", status="surveyed", tags=["第一人称图册"], |
| source="ego", method="pseudo", gt="DFM HandTrack 预测", tier="S5", route="R7", |
| camera="第一视角", rig={"ego": True}, |
| fills="第一视角，模型预测的手部。", |
| team=[T_EGO]) |
| add(id="egolive", name="EgoLive", year="2026", status="surveyed", tags=["第一人称图册"], |
| source="ego", method="pseudo", gt="HaMeR 预测 MANO", tier="S5", route="R7", |
| camera="第一视角", rig={"ego": True}, |
| fills="第一视角，HaMeR 伪标签。", |
| team=[T_EGO]) |
| add(id="egoverse", name="EgoVerse", year="2026", status="surveyed", tags=["第一人称图册"], |
| source="ego", method="device", gt="21 个 3D 关节", tier="-", route="R3", |
| camera="第一视角", rig={"ego": True}, |
| fills="多地采集的第一视角人类示范。", |
| risk="关节来源待核。", |
| links=[arxiv("2604.07607")], team=[T_EGO]) |
| add(id="ipnhands", name="IPN HandS", year="2020", status="surveyed", tags=["2D 图册"], |
| source="exo", method="manual2d", gt="2D 关键点（21 点）", tier="-", route="R4", |
| camera="单相机", rig={"exo": 1}, |
| scale="200 段视频，约 80 万帧", |
| fills="MediaPipe 预标后人工修正的连续 2D。", |
| risk="GT 配对和元数据校验未完成。", |
| team=[T_2D]) |
|
|
| |
| add(id="hograspnet", name="HOGraspNet", year="ECCV 2024", status="p0", tags=["新增"], |
| source="exo", method="rgbd", gt="MANO + HALO", tier="S3", route="R1", |
| camera="4 台 Azure Kinect RGB-D，其中 1 台在身后眼睛高度模拟第一视角", rig={"exo": 4, "depth": True, "pseudoEgo": True}, |
| subjects="99 人，10 到 74 岁", scale="1,489,112 帧，28 类抓取，30 个物体", |
| fills="shape 多样性：作者的 t-SNE 显示比 HO3D、DexYCB、OakInk 更分散。相机与 Femto Bolt 同一传感器族。", |
| risk="手上没有 marker，MANO 来自 MediaPipe 三角化加 RGB-D 拟合。", |
| links=[{"label": "项目主页", "url": "https://hograspnet2024.github.io/"}, arxiv("2409.04033")]) |
| add(id="palm", name="PALM", year="2025", status="p0", tags=["新增"], |
| source="scan", method="scan", gt="扫描网格 + MANO 配准", tier="S1", route="R6", |
| camera="3dMD 扫描仪：7 台 RGB + 14 台黑白", rig={"scanner": True}, |
| subjects="263 人，男女各半，21 到 70 岁", scale="约 9 万张多视角 RGB，1.3 万次扫描，约 50 种右手手势", |
| fills="目前能找到的最可靠 shape 真值，可做 shape 先验和 shape 评测集。", |
| risk="作者称发表后公开；Voxel51 上有 1000 样本子集，完整版获取方式待确认。只有右手静态手势。", |
| links=[arxiv("2511.05403"), {"label": "Voxel51 子集", "url": "https://docs.voxel51.com/dataset_zoo/datasets_hf/palm.html"}]) |
| add(id="egoexo4d", name="Ego-Exo4D 手部姿态", year="CVPR 2024", status="p0", tags=["新增"], |
| source="egoexo", method="mvkp", gt="21 个 3D 关键点", tier="-", route="R3", |
| camera="Aria RGB（110° 视场，卷帘快门）+ 外视角相机", rig={"ego": True, "fisheye": True, "exo": 5}, |
| scale="约 6.8 万人工 2D、34 万人工 3D；自动 3D 约 2100 万", |
| fills="真实场景的宽视场第一视角，对应 HOT3D 上暴露的相机域问题。", |
| risk="没有 MANO，只能监督 pose。", |
| links=[arxiv("2311.18259")]) |
| add(id="gigahands", name="GigaHands", year="CVPR 2025", status="p1", tags=["新增"], |
| source="exo", method="mvkp", gt="MANO", tier="S4", route="R1", |
| camera="51 台 RGB，只有外视角", rig={"exo": 51}, |
| subjects="56 人", scale="34 小时，1.83 亿帧，8.4 万条文本描述", |
| fills="双手动作的多样性。", |
| risk="YOLO + HaMeR + ViTPose 初始化后三角化，再用 EasyMoCap 拟合，可能继承 HaMeR 的偏差，建议先抽检。", |
| links=[{"label": "项目主页", "url": "https://ivl.cs.brown.edu/research/gigahands.html"}, arxiv("2412.04244")]) |
| add(id="oakink2", name="OakInk2", year="CVPR 2024", status="p1", tags=["新增"], |
| source="egoexo", method="marker", gt="SMPL-X → MANO", tier="S2", route="R2", |
| camera="1 个第一视角 + 3 个外视角 RGB，848×480", rig={"ego": True, "exo": 3}, |
| subjects="9 人", scale="627 个序列，401 万帧", |
| fills="第一视角加光学真值的双手复杂任务。", |
| risk="团队 10.8 只用了它的动作分布，图像未入清单。画面中可能看得到 marker，需检查。分辨率低。", |
| links=[{"label": "项目主页", "url": "https://oakink.net/v2"}, arxiv("2403.19417")]) |
| add(id="hint", name="HInt", year="CVPR 2024", status="p1", tags=["新增"], |
| source="wild", method="manual2d", gt="21 点 2D + 逐点遮挡标签", tier="-", route="R4", |
| camera="Ego4D、EPIC VISOR、Hands23、NewDays 图像", rig={"wild": True, "ego": True}, |
| fills="野外第一视角的 2D 监督，以及按遮挡分层的评测。", |
| risk="Ego4D 帧需单独申请许可。", |
| links=[{"label": "GitHub（MIT）", "url": "https://github.com/ddshan/hint"}]) |
| add(id="blurhand", name="BlurHand", year="CVPR 2023", status="p1", tags=["新增"], |
| source="exo", method="mvkp", gt="首、中、尾三帧 3D 真值", tier="S4", route="R1", |
| camera="由 InterHand2.6M 插帧合成运动模糊", rig={"exo": 12}, |
| scale="训练 121,839 张，测试 34,057 张", |
| fills="运动模糊，对应 V7 台本里的快速动作。", |
| links=[{"label": "GitHub", "url": "https://github.com/JaehaKim97/BlurHand_RELEASE"}, arxiv("2303.15417")]) |
| add(id="dartset", name="DARTset", year="NeurIPS 2022", status="p1", tags=["新增"], |
| source="synth", method="synth", gt="MANO 扩展（DART）", tier="SY", route="R1", |
| camera="合成渲染", rig={"synth": True}, |
| scale="约 80 万张，50 种饰品，325 种皮肤纹理", |
| fills="外观多样性：戒指、手表、美甲、纹身。", |
| links=[{"label": "NeurIPS 论文", "url": "https://proceedings.neurips.cc/paper_files/paper/2022/hash/f06d5ebd4ff40b40dd97e30cee632123-Abstract-Datasets_and_Benchmarks.html"}]) |
| add(id="humbi", name="HUMBI Hand", year="CVPR 2020", status="p2", tags=["新增"], |
| source="exo", method="mvkp", gt="MANO", tier="S4", route="R1", |
| camera="手部 68 个视角", rig={"exo": 68}, |
| subjects="全集 772 人，性别均衡，多种肤色", |
| fills="shape 分布广。", |
| risk="每人 shape 取关键点拟合结果的时间中值，只能当弱监督。", |
| links=[{"label": "项目主页", "url": "http://humbi-data.net"}, arxiv("2110.00119")]) |
|
|
| |
| add(id="ourcapture", name="单人手部采集（V6 / V7）", year="2026", status="own", tags=["自采"], |
| source="egoexo", method="marker", gt="SOMA（目标）", tier="S2", route="R0", |
| camera="Femto Bolt 头戴 RGB（带 marker）+ 光学动捕", rig={"ego": True, "exo": 6}, |
| subjects="3 人（计划增加）", scale="V6 152 条 + V7 新增 51 条", |
| fills="极端抓取、手内换抓、手手接触、注视与速度变化。", |
| risk="手上 marker 出现在 RGB 中，需按已知位置做 inpainting。") |
|
|
| |
| add(id="assemblyhands", name="AssemblyHands", year="CVPR 2023", status="no", tags=["不建议"], |
| source="norgb", method="mvkp", gt="21 个 3D 关键点", tier="-", route="R3", |
| camera="第一视角为黑白鱼眼；外视角为 RGB", rig={"ego": True, "fisheye": True, "mono": True, "exo": 8}, |
| subjects="34 人", scale="约 300 万张，其中第一视角约 49 万张", |
| fills="只有外视角 RGB 可用。", |
| risk="第一视角不是真实 RGB，与排除 SHOW3D 的理由相同。", |
| links=[arxiv("2304.12301")]) |
| add(id="show3d", name="SHOW3D", year="2026", status="no", tags=["不建议", "第一人称图册"], |
| source="norgb", method="mvkp", gt="UmeTrack（另提供 MANO / MHR 转换）", tier="S4", route="R1", |
| camera="黑白鱼眼", rig={"ego": True, "fisheye": True, "mono": True}, |
| fills="团队已排除。", |
| risk="无标记多视角自动拟合，人工核验子集不具代表性；黑白鱼眼不满足真实 RGB 条件。", |
| team=[T_EGO]) |
| add(id="hiphi", name="HiPHI", year="2026", status="no", tags=["不建议"], |
| source="norgb", method="marker", gt="BVH 骨架", tier="-", route="R3", |
| camera="无 RGB", rig={"none": True}, |
| subjects="132 人", scale="617.5 小时（含镜像），约 2 亿帧", |
| fills="大规模全身与物体交互动捕，可作动作先验。", |
| risk="没有图像，手部细节未说明。", |
| links=[arxiv("2608.16222")]) |
| add(id="eventegohands", name="EventEgoHands++（EEH-R）", year="2026", status="no", tags=["不建议"], |
| source="norgb", method="device", gt="MANO（手套关节拟合）", tier="S4", route="R1", |
| camera="DAVIS346 事件相机", rig={"ego": True, "none": True}, |
| subjects="8 人", scale="1,019,716 帧，约 2.36 小时", |
| fills="暗光与事件相机。", |
| risk="没有 RGB，公开情况未说明。", |
| links=[{"label": "alphaXiv", "url": "https://www.alphaxiv.org/abs/2609.17189"}]) |
|
|
| out = {"meta": meta, "datasets": D} |
| (ROOT / "data").mkdir(exist_ok=True) |
| (ROOT / "data" / "datasets.json").write_text(json.dumps(out, ensure_ascii=False, indent=1), encoding="utf-8") |
| prev = ROOT / "data" / "previews.json" |
| if not prev.exists(): |
| prev.write_text("{}\n", encoding="utf-8") |
|
|
| |
| ids = [d["id"] for d in D] |
| assert len(ids) == len(set(ids)), "id 重复" |
| src = {s["id"] for s in meta["sources"]} |
| mth = {m["id"] for m in meta["methods"]} |
| tr = {t["id"] for t in meta["tiers"]} |
| st = {s["id"] for s in meta["statuses"]} |
| rt = {r["id"] for r in meta["routes"]} |
| for d in D: |
| assert d["source"] in src, d["id"] |
| assert d["method"] in mth, d["id"] |
| assert d["tier"] in tr, d["id"] |
| assert d["status"] in st, d["id"] |
| assert d["route"] in rt, d["id"] |
| print(f"写出 {len(D)} 个数据集") |
|
|
