Ashadsfasd/mint_ego_datasets

English first — 中文说明在下半部分。

wuji_ego_mint — structured egocentric supervision (non-video)

Camera trajectories, two-hand MANO parameters, per-frame hand presence, and field-of-view labels for 1,021.514 hours of egocentric video, produced by the EgoPipeline data engine of wuji-ego-mint and stored in LeRobot v3.0 layout.

This repository is the pre-training corpus for the MINT model (checkpoint).

Read this first

• No .mp4 files are included. Source video belongs to Ego4D, EgoDex, and EPIC-KITCHENS and cannot be redistributed here. See Obtaining video.
• Labels are pipeline pseudo-labels, not ground truth. They were produced by a monocular reconstruction pipeline, not by motion capture.
• Camera translations exhibit a known scale inflation. Use them for pre-training only. Do not use them for metric-scale evaluation, camera trajectory benchmarking, or as real-scale ground truth.

Contents

Three independent LeRobot v3.0 datasets, one per source corpus:

Subset	Episodes	Frames	Hours @30fps	Stored video shape [H,W,C]	Data chunks
ego4d	190,298	35,950,413	332.874	[768,1024,3]	2
egodex	343,583	68,412,147	633.446	[384,512,3]	4
epickitchen	26,768	5,960,998	55.194	[384,512,3]	1
Total	560,649	110,323,558	1,021.514

Retained after filtering and trajectory cleanup: 59.1 % of 1,729 raw source hours. Total download size is roughly 112 GB across 62,460 files.

Every subset shares one schema: codebase_version: v3.0, robot_type: ego_hand, fps: 30.0, hand_frame: camera, chunks_size: 1000, OpenCV camera convention.

Directory layout

<subset>/lerobot_v3/
|-- meta/
|   |-- info.json                                  dataset-level schema and totals
|   |-- tasks.parquet                              task index -> task text
|   |-- episodes/chunk-NNN/file-NNN.parquet        one row per episode
|   |-- .clip_index.cl32.observation_images_ego.pkl  optional loader cache
|   `-- .clip_index.cl8.observation_images_ego.pkl   optional loader cache
`-- data/chunk-NNN/file-NNN.parquet                one row per frame

<subset> is ego4d, egodex, or epickitchen. There is no videos/ directory. The .clip_index.*.pkl files are caches the MINT loader writes for a given clip_len; they are pickles, they are optional, and the loader regenerates them if absent — delete them if you prefer not to unpickle third-party files.

meta/info.json

Field	Value in this release	Meaning
codebase_version	"v3.0"	LeRobot dataset format version.
robot_type	"ego_hand"	Embodiment tag; these are human hands, not a robot.
fps	30.0	Label and video sampling rate. Converts a timestamp to a frame index.
hand_frame	"camera"	All MANO root translation and orientation is expressed in the current camera frame, not in world space.
chunks_size	1000	Maximum file_index per chunk directory.
total_episodes / total_frames / total_tasks	see table above	Per-subset totals.
data_path	data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet	Template for resolving a frame's Parquet file.
video_path	videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4	Template retained from the original export. Those files are not in this repository.
features	20 entries	Per-column dtype and shape, including observation.images.ego with dtype: "video".

meta/episodes/**/*.parquet — one row per episode

Column	Type	Definition
episode_index	int64	Episode id, unique within the subset.
length	int64	Episode length in frames.
tasks	list<string>	Natural-language task descriptions for the episode.
atomic_segments	list<struct{atomic_index, task_index, start_frame, end_frame}>	Sub-task segmentation. Present in egodex and epickitchen only; absent from ego4d.
data/chunk_index	int64	Chunk of the frame Parquet holding this episode.
data/file_index	int64	File index of that Parquet.
dataset_from_index	int64	Inclusive global row index of the episode's first frame.
dataset_to_index	int64	Exclusive global row index just past the episode's last frame.
meta/episodes/chunk_index	int64	Self-reference: chunk of this episode-metadata file.
meta/episodes/file_index	int64	Self-reference: file index of this episode-metadata file.
videos/observation.images.ego/chunk_index	int64	Chunk of the MP4 in the original export.
videos/observation.images.ego/file_index	int64	File index of that MP4.
videos/observation.images.ego/from_timestamp	float64	Episode start time, in seconds, inside that MP4.
videos/observation.images.ego/to_timestamp	float64	Episode end time, in seconds, inside that MP4.

To locate an episode's frames, read the Parquet named by data/chunk_index and data/file_index, then subtract that file's smallest index value from dataset_from_index to get the local row offset.

An episode is always fully contained in a single frame Parquet file.

data/**/*.parquet — one row per frame

Rows are ordered by episode_index, then frame_index.

Identifiers and timing

Column	Type	Definition
index	int64	Global frame row index within the subset.
episode_index	int64	Owning episode.
frame_index	int64	Zero-based frame number within the episode; starts at 0, increments by exactly 1.
task_index	int64	Row index into meta/tasks.parquet.
main_type	int64	Internal source/category tag. Not used by MINT training.
timestamp	float64	frame_index / fps, in seconds.

Camera

Column	Type	Definition
cam_trans	fixed_size_list<float32>[3]	Translation t of the OpenCV world-to-camera transform [R|t], in meters.
cam_quat	fixed_size_list<float32>[4]	The same transform's rotation R as a unit quaternion in [x, y, z, w] order — real component last.
cam_fov	fixed_size_list<float32>[2]	[fov_from_height, fov_from_width] in radians: 2*atan((H/2)/fy) and 2*atan((W/2)/fx), where (H, W) is the stored video shape.

Camera axes follow OpenCV: +x right, +y down, +z forward. Quaternions are finite and approximately unit length.

The world frame is approximately anchored at each episode's first camera pose: across sampled episodes the median |cam_trans[0]| is 1–2 mm and cam_quat[0].w is ~1. It is not exact — the 95th percentile reaches 1–4 cm and outliers reach ~28 cm. Rebase explicitly rather than assuming the first frame is identity.

The scale-inflation caveat applies to cam_trans. Rotations are unaffected.

Hand presence

Column	Type	Definition
hand_kept	fixed_size_list<bool>[2]	Per-frame [left_valid, right_valid]. This is the authoritative mask: it is both the target of the presence head and the mask gating every MANO loss. false means that side must not contribute a hand loss on that frame.
state_mask	fixed_size_list<bool>[2]	Legacy [left, right] availability flag, retained for LeRobot compatibility. Prefer hand_kept.

Invalid rows still contain finite placeholder MANO values. Always gate on hand_kept.

MANO, per side

{side} is exactly left or right. All values are in the current camera frame (hand_frame: camera).

Column	Type	Definition
{side}_mano_transl_cam	fixed_size_list<float32>[3]	MANO wrist/root translation in meters, camera frame.
{side}_mano_orient6d	fixed_size_list<float32>[6]	Global wrist/root orientation, camera frame, as row-major rotation 6D: the flattened first two rows of the 3x3 rotation matrix. Recover R by Gram-Schmidt.
{side}_mano_pose6d	fixed_size_list<float32>[90]	15 local MANO joint rotations in official MANO order, same row-major 6D convention: 15 * 6 = 90.
{side}_mano_betas	fixed_size_list<float32>[10]	10 dimensionless MANO shape coefficients, stored per frame even when constant across an episode.

MANO mesh assets are not included and are not ours to distribute. Obtain them from mano.is.tue.mpg.de under their own license.

Each side contributes 109 values in this order:

transl_cam[3] + orient6d[6] + pose6d[90] + betas[10] = 109

MINT concatenates left before right into a single target:

hand_gt[218] = left[109] + right[109]

meta/tasks.parquet

Maps task_index to task text. Not read by the MINT training loader; useful for language-conditioned work.

Task strings are bilingual in a single field, formatted 中文:<Chinese>英文:<English>. An episode usually carries several descriptions at different specificity, e.g.:

中文:用蓝色抹布擦拭放有红色水壶的木制台面。英文:Wipe the wooden countertop with the red kettle using the blue cloth.
中文:右手用蓝色抹布擦拭放有红色水壶的木制台面。英文:Wipe the wooden countertop with the red kettle using the blue cloth with the right hand.

Split on the 英文: marker to recover either language.

How MINT consumes this data

Reference implementation: model_train/data/lingbotmap/lerobot_v3.py. Full contract: docs/lerobot-training-data.md.

Clips are non-overlapping and never cross an episode boundary:

clip_len: 32
clip_stride: 32
size_hw: [378, 518]

For an episode of length L, clip starts are 0, 32, 64, ... while start + 32 <= L. A tail shorter than 32 frames is dropped, so episodes shorter than 32 frames yield no samples.

Camera extrinsics are rebased to the first frame of each clip. For stored world-to-camera R[s], t[s]:

R_rel[s] = R[s] @ transpose(R[0])
t_rel[s] = t[s] - R_rel[s] @ t[0]

The first frame is therefore identity rotation and zero translation. cam_fov is not rebased. The camera target is:

gt_pose_enc[S,9] = t_rel[3] + quat_xyzw(R_rel)[4] + cam_fov[2]

Rebasing to a clip-relative frame is what makes the inflated absolute scale tolerable for pre-training — one more reason not to read cam_trans as metric ground truth.

Obtaining video

Frames align to the original export as:

video_frame[k] = round(from_timestamp * fps) + frame_index_offset + k

Because the MP4 files are absent, the videos/observation.images.ego/{chunk_index,file_index,from_timestamp,to_timestamp} columns describe the original export's chunked video layout, which you will not reproduce by downloading source video alone. To train with imagery you must either re-run the repository's LeRobot export over the source corpora, or build your own video files and rewrite those four columns to match.

Apply for source video from its official channels and confirm your own usage and redistribution rights:

• Ego4D
• EgoDex
• EPIC-KITCHENS-100

Loading

import pyarrow.parquet as pq

ep = pq.read_table("epickitchen/lerobot_v3/meta/episodes/chunk-000/file-000.parquet")
row = ep.slice(0, 1).to_pylist()[0]

frames = pq.read_table(
    f"epickitchen/lerobot_v3/data/"
    f"chunk-{row['data/chunk_index']:03d}/file-{row['data/file_index']:03d}.parquet"
)
base = frames.column("index")[0].as_py()
lo = row["dataset_from_index"] - base
episode = frames.slice(lo, row["length"])

print(episode.column("cam_trans")[0], episode.column("hand_kept")[0])

Download the whole corpus with:

hf download ZZJAsher/wuji_ego_mint --repo-type=dataset --local-dir wuji_ego_mint

A single subset:

hf download ZZJAsher/wuji_ego_mint --repo-type=dataset --include "epickitchen/**" --local-dir wuji_ego_mint

Known limitations

• Pipeline pseudo-labels, not ground truth. Do not report MINT-style accuracy against them.
• Camera translation scale is inflated. Pre-training only; no metric evaluation.
• Presence labels inherit a fixed detector threshold from the pipeline.
• No video. The video-locating columns only make sense against the original export.
• MANO assets are not included.

Mirror

Also published on ModelScope: AsherZhu/wuji_ego_mint.

License

The label files in this repository are released under the MIT License, matching wuji-ego-mint's original code. Source video, MANO assets, and upstream models retain their own licenses and access terms; these labels are derived from Ego4D, EgoDex, and EPIC-KITCHENS, whose terms continue to govern your use of the underlying material. Review THIRD_PARTY_NOTICES.md before redistributing anything derived from this release.

Citation

@misc{zhu2026mint,
  title  = {MINT: Minting a Unified Model for World-Space Camera and Hand Motion
            from Scalable Egocentric Pipeline Supervision},
  author = {Zhu, Zijie and Cai, Weiren and Wang, Yizhou and Yang, Zhenjie and
            Liu, Yide and Chen, Jiahao and He, Guanqi},
  year   = {2026},
  note   = {Manuscript under review}
}

Please also cite Ego4D, EgoDex, and EPIC-KITCHENS.

---

wuji_ego_mint —— 结构化第一视角监督数据（非视频部分）

1,021.514 小时第一视角视频对应的相机轨迹、双手 MANO 参数、逐帧手部存在性与 视场角标注，由 wuji-ego-mint 的 EgoPipeline 数据管线产出，以 LeRobot v3.0 格式存储。

本数据集是 MINT 模型（checkpoint）的预训练语料。

使用前必读

• 不包含任何 .mp4 文件。 原始视频版权归 Ego4D、EgoDex、EPIC-KITCHENS 所有， 本项目无权再分发，详见获取视频。
• 标注是管线伪标签，不是真值。 由单目重建管线产出，非动捕采集。
• 相机平移存在已知的尺度放大现象。 请仅用于预训练，不要用于真实尺度评估、 相机轨迹评测，或当作真实尺度 Ground Truth。

数据构成

三个独立的 LeRobot v3.0 数据集，各对应一个来源语料：

子集	Episode 数	帧数	时长（30fps）	原视频尺寸 [H,W,C]	data 分块数
ego4d	190,298	35,950,413	332.874 h	[768,1024,3]	2
egodex	343,583	68,412,147	633.446 h	[384,512,3]	4
epickitchen	26,768	5,960,998	55.194 h	[384,512,3]	1
合计	560,649	110,323,558	1,021.514 h

过滤与轨迹清理后保留 1,729 原始小时的 59.1 %。共 62,460 个文件，下载总量约 112 GB。

三个子集共用同一套 schema：codebase_version: v3.0、robot_type: ego_hand、 fps: 30.0、hand_frame: camera、chunks_size: 1000，相机遵循 OpenCV 约定。

目录结构

<子集>/lerobot_v3/
|-- meta/
|   |-- info.json                                  数据集级 schema 与总量
|   |-- tasks.parquet                              task_index -> 任务文本
|   |-- episodes/chunk-NNN/file-NNN.parquet         每行一个 episode
|   |-- .clip_index.cl32.observation_images_ego.pkl  可选的加载器缓存
|   `-- .clip_index.cl8.observation_images_ego.pkl   可选的加载器缓存
`-- data/chunk-NNN/file-NNN.parquet                每行一帧

<子集> 为 ego4d、egodex 或 epickitchen。没有 videos/ 目录。 .clip_index.*.pkl 是 MINT 加载器针对某个 clip_len 写出的缓存，为 pickle 格式、 可选，缺失时加载器会自动重建 —— 如不愿反序列化第三方 pickle 文件，可直接删除。

meta/info.json

字段	本次取值	含义
codebase_version	"v3.0"	LeRobot 数据格式版本。
robot_type	"ego_hand"	本体标签；这里是人手，不是机器人。
fps	30.0	标注与视频采样率，用于把时间戳换算成帧号。
hand_frame	"camera"	所有 MANO 根节点平移与朝向都在当前相机坐标系下，不是世界坐标系。
chunks_size	1000	每个 chunk 目录内 file_index 的上限。
total_episodes / total_frames / total_tasks	见上表	各子集总量。
data_path	data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet	定位逐帧 Parquet 的模板。
video_path	videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4	沿用原始导出的模板，对应文件不在本仓库内。
features	20 项	各列的 dtype 与 shape，其中 observation.images.ego 的 dtype 为 "video"。

meta/episodes/**/*.parquet —— 每行一个 episode

列名	类型	定义
episode_index	int64	子集内唯一的 episode 标识。
length	int64	episode 帧数。
tasks	list<string>	该 episode 的自然语言任务描述。
atomic_segments	list<struct{atomic_index, task_index, start_frame, end_frame}>	子任务切分。仅 egodex 与 epickitchen 有此列，ego4d 没有。
data/chunk_index	int64	存放该 episode 帧数据的 Parquet 所在 chunk。
data/file_index	int64	该 Parquet 的文件序号。
dataset_from_index	int64	episode 首帧的全局行号（闭区间）。
dataset_to_index	int64	episode 末帧之后的全局行号（开区间）。
meta/episodes/chunk_index	int64	自引用：本 episode 元数据文件所在 chunk。
meta/episodes/file_index	int64	自引用：本 episode 元数据文件序号。
videos/observation.images.ego/chunk_index	int64	原始导出中 MP4 所在 chunk。
videos/observation.images.ego/file_index	int64	该 MP4 的文件序号。
videos/observation.images.ego/from_timestamp	float64	episode 在该 MP4 内的起始秒数。
videos/observation.images.ego/to_timestamp	float64	episode 在该 MP4 内的结束秒数。

定位某个 episode 的帧：按 data/chunk_index 与 data/file_index 打开对应 Parquet，用 dataset_from_index 减去该文件内最小的 index 值，得到局部行偏移。

每个 episode 一定完整落在单个帧 Parquet 文件内。

data/**/*.parquet —— 每行一帧

行序为先 episode_index、后 frame_index。

标识与时间

列名	类型	定义
index	int64	子集内的全局帧行号。
episode_index	int64	所属 episode。
frame_index	int64	episode 内从 0 开始的帧号，必须从 0 起、步长恰为 1。
task_index	int64	指向 meta/tasks.parquet 的行号。
main_type	int64	内部来源/类别标签，MINT 训练不使用。
timestamp	float64	frame_index / fps，单位秒。

相机

列名	类型	定义
cam_trans	fixed_size_list<float32>[3]	OpenCV world-to-camera 变换 [R|t] 中的平移 t，单位米。
cam_quat	fixed_size_list<float32>[4]	同一变换的旋转 R，单位四元数，顺序为 [x, y, z, w]，实部在最后。
cam_fov	fixed_size_list<float32>[2]	[按高计算的 FoV, 按宽计算的 FoV]，单位弧度：2*atan((H/2)/fy) 与 2*atan((W/2)/fx)，其中 (H, W) 为原视频尺寸。

相机轴向遵循 OpenCV：+x 右、+y 下、+z 前。四元数均为有限值且近似单位长度。

世界坐标系近似锚定在每个 episode 的首帧相机位姿上：抽样统计下 |cam_trans[0]| 的中位数为 1–2 mm，cam_quat[0].w 约为 1。但并不精确 —— 95 分位可达 1–4 cm， 个别离群值约 28 cm。请显式重定基，不要假定首帧就是单位位姿。

尺度放大只影响 cam_trans，旋转不受影响。

手部存在性

列名	类型	定义
hand_kept	fixed_size_list<bool>[2]	逐帧 [左手有效, 右手有效]。这是权威掩码：既是存在性分支的监督目标，也是所有 MANO 损失的门控。false 表示该帧该侧不得参与手部损失。
state_mask	fixed_size_list<bool>[2]	旧版 [左, 右] 可用性标记，为兼容 LeRobot 保留。请优先使用 hand_kept。

无效帧仍会写入有限的占位 MANO 数值，务必用 hand_kept 做门控。

逐侧 MANO

{side} 取值严格为 left 或 right。所有数值都在当前相机坐标系下 （hand_frame: camera）。

列名	类型	定义
{side}_mano_transl_cam	fixed_size_list<float32>[3]	MANO 手腕/根节点平移，单位米，相机系。
{side}_mano_orient6d	fixed_size_list<float32>[6]	手腕/根节点全局朝向，相机系，采用行主序 6D 旋转：3x3 旋转矩阵前两行展平。用 Gram-Schmidt 还原 R。
{side}_mano_pose6d	fixed_size_list<float32>[90]	15 个 MANO 局部关节旋转，官方 MANO 顺序，同样的行主序 6D 表示：15 * 6 = 90。
{side}_mano_betas	fixed_size_list<float32>[10]	10 维无量纲 MANO 形状系数。即使在整个 episode 内恒定，也逐帧存储。

不包含 MANO 模型文件，本项目无权分发。请从 mano.is.tue.mpg.de 按其许可条款自行获取。

每侧按此顺序贡献 109 个数值：

transl_cam[3] + orient6d[6] + pose6d[90] + betas[10] = 109

MINT 将左手拼在右手之前，构成单一监督目标：

hand_gt[218] = left[109] + right[109]

meta/tasks.parquet

task_index 到任务文本的映射。MINT 训练加载器不读取此文件，做语言条件相关工作时可用。

任务文本是中英双语写在同一个字段里的，格式为 中文:<中文描述>英文:<English>。 一个 episode 通常带有若干粒度不同的描述，例如：

中文:用蓝色抹布擦拭放有红色水壶的木制台面。英文:Wipe the wooden countertop with the red kettle using the blue cloth.
中文:右手用蓝色抹布擦拭放有红色水壶的木制台面。英文:Wipe the wooden countertop with the red kettle using the blue cloth with the right hand.

以 英文: 为分隔符切分即可取出任一语言。

MINT 如何消费这份数据

参考实现： model_train/data/lingbotmap/lerobot_v3.py。 完整契约： docs/lerobot-training-data.md。

片段互不重叠，且绝不跨越 episode 边界：

clip_len: 32
clip_stride: 32
size_hw: [378, 518]

对长度为 L 的 episode，片段起点为 0, 32, 64, ...，条件是 start + 32 <= L。 不足 32 帧的尾部丢弃，因此短于 32 帧的 episode 不产生任何训练样本。

相机外参会以每个片段的首帧为基准重定基。对存储的 world-to-camera R[s]、t[s]：

R_rel[s] = R[s] @ transpose(R[0])
t_rel[s] = t[s] - R_rel[s] @ t[0]

因此首帧为单位旋转与零平移。cam_fov 不参与重定基。相机监督目标为：

gt_pose_enc[S,9] = t_rel[3] + quat_xyzw(R_rel)[4] + cam_fov[2]

正是这种片段相对化，使得绝对尺度的放大在预训练中可以容忍 —— 这也是不应把 cam_trans 当作米制真值的又一个原因。

获取视频

帧与原始导出的对齐关系为：

video_frame[k] = round(from_timestamp * fps) + 片段内偏移 + k

由于 MP4 未包含在内， videos/observation.images.ego/{chunk_index,file_index,from_timestamp,to_timestamp} 这四列描述的是原始导出的分块视频布局 —— 仅靠下载原始视频无法复现该布局。 若要带图像训练，需要二选一：用本仓库的 LeRobot 导出流程在原始语料上重跑一遍， 或者自行构建视频文件并改写这四列使其匹配。

请从各数据集官方渠道申请原始视频，并自行确认使用与再分发权限：

• Ego4D
• EgoDex
• EPIC-KITCHENS-100

读取示例

import pyarrow.parquet as pq

ep = pq.read_table("epickitchen/lerobot_v3/meta/episodes/chunk-000/file-000.parquet")
row = ep.slice(0, 1).to_pylist()[0]

frames = pq.read_table(
    f"epickitchen/lerobot_v3/data/"
    f"chunk-{row['data/chunk_index']:03d}/file-{row['data/file_index']:03d}.parquet"
)
base = frames.column("index")[0].as_py()
lo = row["dataset_from_index"] - base
episode = frames.slice(lo, row["length"])

print(episode.column("cam_trans")[0], episode.column("hand_kept")[0])

下载全量：

hf download ZZJAsher/wuji_ego_mint --repo-type=dataset --local-dir wuji_ego_mint

只下载单个子集：

hf download ZZJAsher/wuji_ego_mint --repo-type=dataset --include "epickitchen/**" --local-dir wuji_ego_mint

已知局限

• 管线伪标签，不是真值，不要拿它去衡量 MINT 级别的精度。
• 相机平移尺度放大，仅供预训练，不做米制评测。
• 存在性标签继承了管线中固定的检测器阈值。
• 不含视频，视频定位列只在原始导出语境下成立。
• 不含 MANO 模型资产。

镜像

同时发布于 ModelScope： AsherZhu/wuji_ego_mint。

许可

本仓库中的标注文件以 MIT License 发布，与 wuji-ego-mint 的原创代码一致。原始视频、 MANO 资产与上游模型各自保留其许可与访问条款；这些标注派生自 Ego4D、EgoDex 和 EPIC-KITCHENS，你对底层素材的使用仍受其条款约束。再分发本数据的任何派生物之前， 请先阅读 THIRD_PARTY_NOTICES.md。

引用

@misc{zhu2026mint,
  title  = {MINT: Minting a Unified Model for World-Space Camera and Hand Motion
            from Scalable Egocentric Pipeline Supervision},
  author = {Zhu, Zijie and Cai, Weiren and Wang, Yizhou and Yang, Zhenjie and
            Liu, Yide and Chen, Jiahao and He, Guanqi},
  year   = {2026},
  note   = {Manuscript under review}
}

同时请引用 Ego4D、EgoDex 和 EPIC-KITCHENS。