From declarationto download.
Ampe-DB is openly available to everyone and free of charge. Before downloading, each user completes a short access form stating their intended purpose. It is a required use declaration — not an approval process — and there is no waiting period.
Access the dataset
Access process
- 01Complete the official access form with your name, contact email, affiliation where applicable, intended use and a short description of the project or purpose.
- 02Confirm acceptance of the Terms of Use, License and Research Data Use Agreement.
- 03Submit the form. A personal download link is emailed to you straight away — no separate approval is required.
Who may download
- Anyone, after completing the access form and agreeing to the release terms.
- Academic, educational, independent, cultural, artistic and exploratory uses are welcome when they comply with the Terms of Use.
- There is no download fee, and institutional affiliation is not required.
Open download access does not waive the Terms of Use. Restrictions on redistribution, re-identification, surveillance, prohibited commercial use and misuse of participant-derived data still apply.
Release packages
Download links are emailed after the declaration| Package | Contents | Archive |
|---|---|---|
| Raw RGB videos | 637 canonical MP4 clips with synchronized embedded audio | videos.zip |
| Normalized pose | Two-player, 33-landmark normalized pose CSV for model inputs | pose.zip |
| Skeleton videos | Rendered skeleton MP4 files for visual inspection | skeleton-videos.zip |
| Skeleton coordinates | Pixel-space landmark exports in CSV and JSON | skeleton-coordinates.zip |
| Annotations & metadata | Movement labels, manifest, participant–pair mapping and schemas | annotations-metadata.zip |
| Splits & documentation | Frozen evaluation partitions, paper, README, citation file and release documentation | research-docs.zip |
What you receive
Approved V1 components
- Canonical RGB MP4 clips with synchronized embedded audio.
- Normalized 2D pose CSV files.
- Skeleton coordinate exports in CSV and JSON.
- Rendered skeleton MP4 visualizations.
- Canonical movement labels and participant/pair mapping metadata.
- Duplicate provenance and media metadata.
- Locked split definitions, release documentation, paper materials, citation file, and access and licensing documents.
Download by research need
| Use case | Minimum useful components |
|---|---|
| RGB action recognition | videosannotationsmetadatasplits |
| Pose sequence modeling | poseannotationsmetadatasplits |
| Skeleton inspection / visualization | skeletonvideosmetadata |
| Benchmark reproduction | required modalityannotationsmetadatasplitsbenchmark documentation |
Download instructions
After completing the access form, open the personal link we email you and download Ampe-DB free of charge. The link works for 7 days; if it expires, submit the form again for a new one. No separate approval is required.
Preserve the directory structure: filenames and relative paths are used to align representations and metadata.
After download
- 01Keep canonical filenames unchanged.
- 02Confirm the expected top-level folders and metadata files are present.
- 03Check canonical identifier coverage before analysis.
- 04Verify package-level checksums when checksum material is supplied with the release.
- 05Record the dataset version in your experiment configuration.
Do not use earlier draft checksum lists as authoritative V1 integrity records. Use only checksum material distributed with the release package you actually downloaded.
Verify download & checksums
Two levels of verification
| Level | What to verify |
|---|---|
| Release / package integrity | Use the checksum file supplied with the specific download package, when provided, to confirm that transferred archives or files are complete. |
| Dataset record integrity | Use metadata/ampe_db_v1_manifest.csv to confirm canonical clip IDs, media properties and the complete-file RGB/audio SHA-256 values used for duplicate provenance. |
Canonical coverage checks
- Each of the 637 canonical IDs should be represented in all five record-level representations: RGB video, normalized pose CSV, skeleton CSV, skeleton JSON and skeleton video.
- A matching movement token +
C###identifier denotes the same canonical record across representations. - The 3,185 count excludes release documentation, metadata, split, paper and access files.
Duplicate provenance
V1 contains 557 unique complete-file RGB/audio SHA-256 hashes across 637 canonical records. Exact duplicates are retained for provenance and record correspondence; duplicate-safe splitting prevents them from crossing the fixed train, validation and test partitions. Dataset overview →
Version information
Version reporting
- Always name the dataset version used.
- State the exact evaluation protocol used.
- Do not compare fixed-split results directly with pair-held-out or participant-held-out results without identifying the protocol.
- Documentation-only corrections should not silently change frozen benchmark definitions or historical results.