Technicaldocumentation.

Dataset overview

Ampe-DB captures two-person interactions from Ampe, a Ghanaian game involving coordinated jumping, clapping and a leg-position reveal. The benchmark treats each segmented clip as a paired movement configuration rather than a solo action.

The dataset is deliberately compact rather than large-scale. Its contribution is the combination of culturally grounded interaction, explicit participant/pair metadata, aligned RGB and pose representations, and leakage-aware generalization protocols.

Movement labels

Each record carries one of four movement classes — left_left, left_right, right_left or right_right — and a binary relation label derived from it. Label spaces →

Database structure

The publication package is organized around eight core folders. The root also holds the release README, citation file, research data use agreement and license.

Canonical 1:1 record representations

Each canonical clip identifier maps to five record-level representations. Across 637 records this produces 3,185 record-level files, before documentation, metadata, split, paper and access files are counted.

Core metadata files

• metadata/ampe_db_v1_manifest.csv
• metadata/ampe_db_v1_schema.csv
• metadata/ampe_db_v1_data_dictionary.csv
• metadata/participants.csv
• metadata/participant_pairs.csv
• metadata/clip_participant_mapping.csv
• metadata/dataset_statistics.csv
• metadata/movement_distribution_summary.csv

Data formats

RGB video and audio

Canonical media are MP4 clips with synchronized embedded audio. Frame rate and codec vary across canonical media and are documented in metadata.

Normalized pose CSV

The canonical pose representation is long-form, with up to 33 MediaPipe landmarks per player per frame:

• x_norm and y_norm are body-normalized coordinates, not simple [0, 1] image coordinates.
• Normalization uses the hip midpoint as origin and a shoulder-to-hip torso scale.
• No rotation or reflection normalization is applied.
• Missing detections and incomplete player-frame groups are handled explicitly during model preparation using presence indicators.

Skeleton exports

Skeleton CSV and JSON files preserve pixel-space landmark geometry for visualization and reconstruction. Skeleton MP4 files are derivative visualizations for inspection, debugging, demonstration and publication figures — not a separate ground-truth modality.

File naming rules

The movement token and C### identifier form the canonical record key used across all representations.

Movement tokens

Pose & skeleton format

Ampe-DB V1 uses the standard 33-landmark MediaPipe pose representation for each represented player.

The release preserves the full landmark index rather than the 21-keypoint subset described in an earlier draft.

Pose CSV schema

A complete frame with two detected players can contain up to 66 landmark rows; missing detections mean some frame-player groups may be incomplete.

Normalization

Baseline V1 model-ready sequence

Landmark index

The standard MediaPipe pose landmarks, as stored in landmark_index. Marked: the shoulder and hip landmarks that define the torso used for normalization.

• 0nose
• 1left_eye_inner
• 2left_eye
• 3left_eye_outer
• 4right_eye_inner
• 5right_eye
• 6right_eye_outer
• 7left_ear
• 8right_ear
• 9mouth_left
• 10mouth_right
• 11left_shoulder
• 12right_shoulder
• 13left_elbow
• 14right_elbow
• 15left_wrist
• 16right_wrist
• 17left_pinky
• 18right_pinky
• 19left_index
• 20right_index
• 21left_thumb
• 22right_thumb
• 23left_hip
• 24right_hip
• 25left_knee
• 26right_knee
• 27left_ankle
• 28right_ankle
• 29left_heel
• 30right_heel
• 31left_foot_index
• 32right_foot_index

Skeleton outputs

Skeleton CSV and JSON expose the pixel-coordinate geometry used for rendered skeleton visualization. Skeleton videos are convenient inspection layers, not substitutes for the original RGB clip or the canonical pose CSV.

Annotation & metadata

annotations/ampe_db_v1_labels.csv stores the V1 clip-level supervised labels: a four-class movement label and a binary movement relation derived from it.

Four classes (cells) and the binary relation they map to: the diagonal is the same leg configuration, the off-diagonal a different one.

Participant and pair mappings

clip_participant_mapping.csv links each canonical clip to pseudonymous player IDs, ordered and unordered pair identifiers, and movement fields. Participant identity is distributed separately from the movement label.

Manifest and duplicate provenance

The 637-row canonical manifest records media properties and complete-file RGB/audio SHA-256 values used to identify exact-duplicate groups.

Participants & pairs

Files to use

Why pair information matters

Pair identity is essential to the V1 evaluation design. The primary generalization benchmark holds out participant pairs as pairs, while individual participants may still appear with different partners in training for a given fold.

Recording conditions

Capture design

• Single fixed frontal camera viewpoint.
• Two players visible together within the frame.
• Source recordings captured natural Ampe play and were later segmented into canonical interactions.
• Embedded audio retains clap, step and vocal cues from the source recording.

Camera equipment

Recording documentation identifies Blackmagic and Sony A7 III cameras in the source capture workflow. Rely on per-file media metadata rather than assuming one uniform camera profile for every canonical clip.

Media properties

V1 media properties are not completely uniform. Read the release metadata instead of hard-coding 60 FPS or a single codec.

Why the frontal view matters

The fixed frontal view keeps both players visible, supports consistent 2D pose extraction, and preserves the relative timing and leg-position structure of Ampe without a multi-view motion-capture rig.

Release manifest & file index

metadata/ampe_db_v1_manifest.csv is the central 637-row record inventory. Use it to join canonical clip identifiers, media properties, duplicate provenance and release-level record information.

Record-level coverage

Supporting index files

• ampe_db_v1_schema.csv
• ampe_db_v1_data_dictionary.csv
• participants.csv
• participant_pairs.csv
• clip_participant_mapping.csv
• dataset_statistics.csv
• movement_distribution_summary.csv