README.md

---
license: cc-by-nc-4.0
pretty_name: Eyes, JAPAN MoCap (Raw Files + MP4 Previews) [Test]
annotations_creators:
- machine-generated
language:
- en
task_categories:
- other
tags:
- motion-capture
- mocap
- human-motion
- animation
- 3d
- video
- tabular
- c3d
- bvh
- fbx
- vicon
- motion-generation
size_categories:
- 10K<n<100K
configs:
- config_name: files
  default: true
  builder_name: videofolder
  data_files:
  - split: train
    path: "metadata.parquet"

- config_name: videos
  builder_name: videofolder
  data_dir: "Videos"
---

# Eyes, JAPAN MoCap Dataset
## Raw Motion-Capture Files and MP4 Preview Videos

> **Temporary test dataset card**
>
> This README reflects the current prepared dataset and its working Hugging Face
> Data Studio configuration. Licensing authority, performer consent, final
> attribution wording, repository identifiers, and citation details should be
> reviewed before the final public release.

This repository contains a standardized, multi-format human motion-capture
collection associated with Eyes, JAPAN. It includes searchable file-level
metadata and a separate set of browser-compatible MP4 preview videos.

The repository exposes two Hugging Face configurations:

| Configuration | Rows | Purpose |
|---|---:|---|
| `files` | 10,254 | One metadata row for each preserved original physical file |
| `videos` | 1,002 | One playable MP4 preview linked to a source C3D motion |

The MP4 previews are derived convenience copies and are not included in the
10,254-file original dataset count.

## Dataset summary

- **Original physical files:** 10,254
- **Logical motion groups:** 4,131
- **Motion categories:** 44
- **C3D files:** 3,962
- **Playable MP4 previews:** 1,002
- **Actor-label values:** 48, including `NA`
- **Top-level source bundles:** `free` and `premium`
- **File ID range:** `00001` to `10254`

Every physical file has a unique zero-padded five-digit `file_id`.

Files representing the same logical motion share a common `motion_name`,
description, and tag set.

## Repository structure

```text
.
├── README.md
├── metadata.parquet
├── free/
│   └── <category>/
│       └── <motion_name>/
│           └── <file_id>-<motion_name>.<extension>
├── premium/
│   └── <category>/
│       └── <motion_name>/
│           └── <file_id>-<motion_name>.<extension>
└── Videos/
    ├── metadata.parquet
    ├── <video_file_id>-<motion_name>.mp4
    └── ...
```

Example original motion group:

```text
premium/
└── badminton/
    └── premium-badminton-ralley1-NA/
        ├── 04990-premium-badminton-ralley1-NA.bip
        ├── 04991-premium-badminton-ralley1-NA.c3d
        ├── 04992-premium-badminton-ralley1-NA.csm
        ├── 04993-premium-badminton-ralley1-NA.v
        └── 04994-premium-badminton-ralley1-NA.mov
```

Corresponding derived preview:

```text
Videos/
└── 04994-premium-badminton-ralley1-NA.mp4
```

The original MOV or M4V file remains unchanged in the source hierarchy. Only
the derived copy in `Videos/` is converted to MP4.

## Standardized naming

Original files follow:

```text
<file_id>-<motion_name>.<extension>
```

The logical motion name generally follows:

```text
<bundle>-<category>-<action>-<actor>
```

Example:

```text
04991-premium-badminton-ralley1-NA.c3d
```

A separately maintained processed NPZ companion dataset keeps the same
five-digit ID as its source C3D:

```text
04991-premium-badminton-ralley1-NA.c3d
04991-premium-badminton-ralley1-NA.npz
```

This shared ID allows direct matching between the original and processed
repositories.

## Original file formats

| Extension | Count | General role |
|---|---:|---|
| `.c3d` | 3,962 | Optical marker and motion-capture data |
| `.bip` | 1,604 | 3ds Max Character Studio Biped motion |
| `.csm` | 1,067 | Character Studio marker data |
| `.v` | 886 | Vicon-related capture data |
| `.mov` | 886 | Original reference video |
| `.bvh` | 788 | Skeletal hierarchy and animation data |
| `.fbx` | 690 | Autodesk FBX animation data |
| `.vsk` | 255 | Vicon skeleton calibration data |
| `.m4v` | 116 | Original reference video |

Some formats may require proprietary or format-specific software for complete
inspection or conversion.

## Configuration: `files`

The default `files` configuration loads the root `metadata.parquet`.

It contains one row per preserved original physical file.

| Column | Description |
|---|---|
| `file_id` | Five-character string matching the filename prefix |
| `bundle` | `free` or `premium` |
| `category` | Motion category |
| `motion_name` | Standardized logical motion identifier |
| `actor_names` | Actor or performer label when available |
| `file_path` | Repository-relative path to the physical file |
| `extension` | Normalized extension without a leading period |
| `size_bytes` | Physical file size in bytes |
| `preview_video_path` | Original MOV/M4V path associated with a C3D row |
| `frame_count` | C3D point-frame count |
| `frame_rate_Hz` | C3D point-data rate |
| `duration_seconds` | C3D duration rounded to one decimal place |
| `description` | Filename-grounded human-readable motion summary |
| `tags` | Semicolon-separated search terms |

Timing fields are populated only for C3D rows. Other formats contain null
timing values.

The `preview_video_path` field is a repository-relative cross-reference to the
original reference video. It is not itself a playable Data Studio video field.

## Configuration: `videos`

The `videos` configuration loads `Videos/` through Hugging Face VideoFolder.

`Videos/metadata.parquet` contains:

| Column | Description |
|---|---|
| `file_name` | MP4 filename relative to `Videos/metadata.parquet` |
| `video_file_id` | ID inherited from the original MOV/M4V |
| `source_c3d_file_id` | ID of the linked source C3D |
| `bundle` | `free` or `premium` |
| `category` | Motion category |
| `motion_name` | Standardized logical motion identifier |
| `actor_names` | Actor or performer label |
| `source_c3d_path` | Repository-relative source C3D path |
| `description` | Motion description copied from the source C3D metadata |
| `tags` | Search tags copied from the source C3D metadata |

VideoFolder interprets `file_name` as the native video feature displayed by
Hugging Face Data Studio.

All preview copies use:

```text
Container:      MP4
Video codec:    H.264
Pixel format:   yuv420p
Fast start:     enabled
Audio:          removed
Frame rate:     inherited from source
Resolution:     preserved, with odd dimensions adjusted when required
```

## Loading the dataset

Replace `<OWNER>/<DATASET_NAME>` with the final repository ID.

### Load the complete file index

```python
from datasets import load_dataset

repo_id = "<OWNER>/<DATASET_NAME>"

files = load_dataset(
    repo_id,
    "files",
    split="train",
)

print(files)
print(files.features)
```

### Load playable preview videos

```python
from datasets import load_dataset

repo_id = "<OWNER>/<DATASET_NAME>"

videos = load_dataset(
    repo_id,
    "videos",
    split="train",
)

print(videos)
print(videos.features)
```

### Filter the file index

```python
c3d_files = files.filter(
    lambda row: row["extension"] == "c3d"
)

walk_c3d = files.filter(
    lambda row:
        row["category"] == "walk"
        and row["extension"] == "c3d"
)
```

### Download a physical file

```python
from huggingface_hub import hf_hub_download

row = files[0]

local_path = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    filename=row["file_path"],
)

print(local_path)
```

### Link a preview to its source C3D

```python
preview = videos[0]

print(preview["video_file_id"])
print(preview["source_c3d_file_id"])
print(preview["source_c3d_path"])
```

## Descriptions and tags

Descriptions and tags are assigned once per unique `motion_name` and shared by
all physical formats representing the same logical motion.

They are generated from available metadata, including:

- bundle;
- category;
- standardized motion name;
- actor label;
- exact hyphen-separated action tokens.

Descriptions may include mechanics that are inherent to a named movement. For
example, a handspring may be described as involving hand support and an
inverted flight phase.

The underlying C3D and video content was not manually inspected for every
entry. Descriptions should therefore be treated as **filename-grounded
metadata annotations**, not expert-verified motion transcriptions.

## Data quality and known anomalies

The metadata was validated for file IDs, paths, extensions, sizes, C3D timing,
video mappings, and Parquet types.

Known retained file anomalies:

### Zero-frame C3D files

Sixteen small C3D files in the premium drum-motion collection report:

```text
frame_count:       0
frame_rate_Hz:     120
duration_seconds:  0.0
size_bytes:        3072
```

These may be empty placeholders or header-only C3D files. They remain indexed
because the metadata accurately reflects the physical files.

### Zero-byte BIP files

Two BIP files report:

```text
size_bytes: 0
```

Affected IDs:

```text
02558
08503
```

These files may be empty or unusable and should be treated cautiously.

### General format considerations

Not every source format has been manually opened and verified in its original
authoring application. Users should independently validate files required for
critical workflows.

## Intended uses

Potential non-commercial uses include:

- motion retrieval and classification;
- human-motion generation;
- animation and character-control research;
- motion-style analysis;
- format-conversion research;
- biomechanics and motion-processing experiments;
- educational demonstrations.

Users should independently verify coordinate systems, units, skeleton
definitions, marker conventions, and format-specific compatibility before
using the data in production pipelines.

## Licensing

This test card currently declares the intended license as
[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/).

Under that license, material may be shared and adapted for non-commercial
purposes when appropriate attribution is provided, the license is linked, and
changes are indicated.

Commercial use is not granted by CC BY-NC 4.0.

**Before final publication, confirm:**

- the rights holder is authorized to license every included file;
- whether all original and derived files are covered by the same terms;
- whether gated access is required;
- the required attribution statement;
- the commercial licensing procedure;
- whether performer privacy, publicity, and consent rights permit release of
  the reference videos.

Commercial licensing and dataset questions:

```text
info@nowhere.co.jp
```

## Data collection and consent

Detailed collection documentation is still under review.

Before final release, add verified information about:

- capture studio and location;
- capture hardware and software;
- marker set and calibration;
- coordinate system and physical units;
- recording frame rates;
- performer recruitment and consent;
- whether consent covers public redistribution of reference video;
- anonymization or exclusion rules;
- known demographic or performance biases.

## Citation

Replace the repository URL and confirm the preferred creator attribution
before final publication.

```bibtex
@misc{eyesjapan_mocap_2026,
  title        = {Eyes, JAPAN MoCap Dataset: Raw Files and MP4 Previews},
  author       = {{Eyes, JAPAN}},
  year         = {2026},
  howpublished = {Hugging Face Dataset},
  url          = {https://huggingface.co/datasets/<OWNER>/<DATASET_NAME>}
}
```

## Final-release checklist

Before removing the `[Test]` label:

- [ ] confirm the final repository ID;
- [ ] confirm the license with the rights holder;
- [ ] confirm reference-video consent and privacy review;
- [ ] decide whether access should be public or gated;
- [ ] verify all 10,254 original paths on the Hub;
- [ ] verify all 1,002 MP4 preview rows in Data Studio;
- [ ] verify the `files` and `videos` configuration selector;
- [ ] verify the zero-frame and zero-byte anomaly documentation;
- [ ] replace placeholder usage URLs;
- [ ] approve the final citation and attribution text;
- [ ] add verified capture-system and coordinate-system documentation.

### Optional gated-access block

The test README intentionally does not enable gated access. After legal review,
a gated-access block can be added to the YAML header:

```yaml
extra_gated_prompt: >-
  This dataset is provided for non-commercial use under the stated license.
  By requesting access, you confirm that you have reviewed and agree to the
  dataset terms.
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
  I agree to use the dataset only as permitted by the license: checkbox
```

## Takedown and corrections

To report a rights concern, metadata error, broken file, or mapping problem,
contact:

```text
info@nowhere.co.jp
```

For metadata corrections, include the five-digit `file_id`, repository path,
and a brief explanation.