01Source
RGB video + audio
Canonical MP4 clips with synchronized embedded audio, retaining clap, step and vocal cues.
videos/Movement is data.Culture is context.
Ampe-DB is a multimodal research dataset of two-person interactions from Ampe, a traditional Ghanaian game. It is built for human activity recognition, pose research and rigorous cross-person generalization studies.
RGB · videos/
Skeleton · skeleton/videos/
637
Canonical two-person clips
15
Pseudonymous participants
11
Unordered player pairs
4
Movement classes
5
Aligned representations
Rooted in living tradition
What is Ampe?
Ampe is a traditional Ghanaian game played by two people who jump, clap and reveal a leg position in a tightly coordinated interaction. Its patterns make it a valuable setting for studying paired movement rather than isolated single-person actions.
Ampe-DB turns short, carefully organized segments of natural gameplay into a resource for computer vision and human interaction research, while preserving the cultural context of the game. It is deliberately compact rather than large-scale.
left_right_C114 · different legs
right_right_C189 · same legs
right_left_C082 · different legs
left_left_C009 · same legs
- 01Jump. Both players jump in a tightly coordinated rhythm.
- 02Clap. Claps punctuate the cycle; the cue survives in each clip’s embedded audio.
- 03Leg reveal. Each player reveals a leg. The pair’s configuration is the V1 label.
What’s in the release
One interaction.Five representations.
Every canonical record identifier maps to five aligned resources. Across 637 records that is 3,185 record-level files, before documentation, metadata, splits and paper materials.
See the folder structure →02Model input
Normalized 2D pose
Long-form CSV with up to 33 MediaPipe landmarks per player per frame, hip-centered and torso-scaled.
pose/03Geometry
Skeleton CSV
Pixel-space landmark coordinates for visualization and reconstruction.
skeleton/csv/04Geometry
Skeleton JSON
A JSON mirror of the same pixel-space skeleton coordinates.
skeleton/json/05Visualization
Skeleton MP4
Rendered videos for inspection and figures — not a separate ground-truth modality.
skeleton/videos/Baseline V1 benchmark
Testing generalization,not memorization.
The central question: do models learn transferable Ampe interaction patterns, or rely on familiarity with the same recordings, player pairings and participants?
Pose-based models transfer substantially better than frozen RGB features once evaluation moves from the duplicate-safe fixed split to unseen player pairs and held-out participants.
−0.537
RGB — four-class Macro-F1, fixed split → unseen pairs
−0.146
Pose — the same shift, far smaller loss
Four-class Macro-F1 by protocol
RGBPoseRGB + Pose
Duplicate-safe fixedReference split
RGB0.955Pose0.957RGB + Pose0.976
Pair-held-outPrimary benchmark · 4 folds
RGB0.418Pose0.811RGB + Pose0.824
Participant-held-outDiagnostic
RGB0.384Pose0.839RGB + Pose0.833
Getting started
Six steps for new users.
From the terms to a reproducible, citable result. Each step links to the page that covers it in full.
- 01
Review the terms
Read the Terms of Use and the Research Data Use Agreement.
Terms of Use → - 02
Declare, then download
Complete the access form stating your intended purpose, then download the dataset free of charge.
Access the dataset → - 03
Take only what you need
Download only the package components your research requires.
What you receive → - 04
Read before writing loaders
Read Database Structure, Data Formats and File Naming Rules first.
Database structure → - 05
Report on the provided splits
Use the provided split definitions when reporting Baseline V1 results.
Evaluation protocols → - 06
Cite version and protocol
Cite Ampe-DB and state the release version and evaluation protocol used.
Citation →
Recommended first files
Open these before writing any code.
| File | Why start here |
|---|---|
README.md | Release-level orientation |
CITATION.cff | Citation metadata |
metadata/ampe_db_v1_manifest.csv | Canonical record inventory and duplicate provenance |
annotations/ampe_db_v1_labels.csv | The two V1 supervised label spaces |
splits/ | Locked fixed and pair-held-out partition files |
News & updates
Release notes.
Entries name the dataset version affected and separate documentation corrections from changes to distributed data or benchmark definitions. Historical benchmark results stay reproducible after documentation updates.
- 2026Ampe-DB V1Documentation
Website documentation aligned to the frozen release structure and the Baseline V1 benchmark.
- 2026Baseline V1Benchmark docs
Baseline V1 documentation standardized around the duplicate-safe fixed reference split, the 4-fold pair-held-out benchmark and the participant-held-out diagnostic.
- 2026Ampe-DB V1Documentation
Documentation updated for the 33-landmark pose representation, non-uniform media properties, duplicate provenance and final RGB, Pose and Fusion benchmark results.
Citation & acknowledgements
Cite Ampe-DB.
The CITATION.cff file at the release root is the source of truth. Do not invent a DOI, repository identifier or publisher field that has not been assigned in the release.
- Authors
- Ruth Senior BineyIsaac Kojo Yedonu Aboah
- Year
- 2026
Suggested plain-text citation
Biney, R. S., & Aboah, I. K. Y. (2026). Ampe-DB: A dataset for human activity and interaction analysis using the Ghanaian game Ampe (Version 1.0.0) [Data set].
Suggested acknowledgement
“This work uses Ampe-DB, created by Ruth Senior Biney and presented at the Young Africans Research Academy (YARA) to support research on human activity and interaction analysis using the Ghanaian game Ampe.”