# **Ampe-DB V1 — Annotation Guide** **Dataset:** Ampe-DB V1 **Documentation Revision:** October 2026 **Research Authors:** Ruth Senior Biney and Isaac Kojo Yedonu Aboah ## **1\. Purpose** This document defines the semantic annotations that may legitimately be treated as Ampe-DB V1 ground truth. It explains the verified movement labels, deterministic derived annotations, pseudonymous participant information, technical quality indicators, and historical annotation products that were audited but not accepted as canonical ground truth. The guiding principle is: **Only labels whose provenance and meaning have been verified are included as canonical V1 ground truth.** Automatically generated annotations must not be presented as human-verified measurements unless their provenance and validation support that claim. ## **2\. Canonical Interaction Records** Each canonical V1 record represents one segmented dyadic interaction clip derived from the Ghanaian game Ampe. The canonical record identifier is `clip_id`. | Characteristic | Count | | :---- | :---- | | Canonical interaction records | 637 | | Unique complete-file RGB/audio SHA-256 hashes | 557 | The canonical annotation file is: `annotations/ampe_db_v1_labels.csv` It contains one row for each of the 637 canonical interaction records. Exact duplicate records are retained as separate canonical records for provenance. However, all records with identical underlying RGB/audio content must remain within the same benchmark partition. Some historical benchmark files use `clip_id_canonical` rather than `clip_id`. These fields refer to the same canonical record concept and should be reconciled carefully when joining the benchmark files with the release manifest. ## **3\. Primary Annotation: Movement Class** The primary verified semantic annotation is: `movement_class` It contains four allowed values: | Class | Canonical records | | :---- | :---- | | `left_left` | 200 | | `left_right` | 134 | | `right_left` | 114 | | `right_right` | 189 | | **Total** | **637** | The labels describe the stored two-token movement configuration observed in each interaction clip. ### **3.1 Interpretation** The safest verified interpretation is relational. | Class | Verified interpretation | | :---- | :---- | | `left_left` | Both movement tokens have the value left | | `right_right` | Both movement tokens have the value right | | `left_right` | The movement tokens differ, with the stored order left then right | | `right_left` | The movement tokens differ, with the stored order right then left | These labels are computational descriptors of movement configurations. They do not encode the complete rules or cultural meaning of Ampe. ### **3.2 Critical Token-Ordering Limitation** The V1 audit did not independently establish that the first and second movement tokens correspond to the canonical `player_1` and `player_2` identities. Consequently, the following interpretation is not independently verified: `left_right` \= `player_1` left and `player_2` right. Similarly, the following is not independently verified: `right_left` \= `player_1` right and `player_2` left. The four-class movement label itself is verified. Its token-to-participant role mapping is not. Models may use the four-class target, but publications and software should not assign unsupported participant-specific meanings to the token positions. ## **4\. Derived Annotation: Movement Relation** The binary movement-relation target is: `movement_relation` It is deterministically derived from the verified four-class movement configuration. | Movement class | Derived movement relation | | :---- | :---- | | `left_left` | `same_leg_configuration` | | `right_right` | `same_leg_configuration` | | `left_right` | `different_leg_configuration` | | `right_left` | `different_leg_configuration` | The derived class distribution is: | Movement relation | Records | | :---- | :---- | | `same_leg_configuration` | 389 | | `different_leg_configuration` | 248 | | **Total** | **637** | The binary movement relation is not an independently annotated label. It is a deterministic transformation of `movement_class`. This distinction must be preserved in research papers, model descriptions, and dataset documentation. ## **5\. Pseudonymous Participant Identity** Ampe-DB V1 contains 15 pseudonymously represented participants: `P01`–`P15` Canonical participant assignments were recovered and manually reviewed during the identity-recovery stage. All 637 interaction records have canonical participant assignments in the release metadata. These participant identifiers support reproducible participant-aware and pair-aware evaluation. Real-world participant names and the internal identity-recovery mapping are excluded from the intended release. Pseudonymous participant identity is a verified record-level annotation, but it remains participant-linked information and is subject to applicable privacy and distribution requirements. ## **6\. Participant-Pair Annotations** Each canonical record includes participant-pair information. The V1 collection contains: | Pair representation | Count | | :---- | :---- | | Ordered participant-pair configurations | 14 | | Unordered participant pairs | 11 | The difference arises because some pairs appear in both Player 1 / Player 2 orders. Participant-pair membership is important for scientific evaluation because pair composition and movement-class distribution are strongly associated. Several participant pairs do not contain examples from every movement class. Researchers should therefore account for participant-pair structure when constructing and interpreting benchmarks. ## **7\. Participant Metadata Files** The canonical release includes three dedicated participant-metadata files. | File | Content | | :---- | :---- | | `metadata/participants.csv` | One record for each of the 15 pseudonymous participants | | `metadata/participant_pairs.csv` | One record for each of the 11 unordered participant pairs | | `metadata/clip_participant_mapping.csv` | Clip-level participant and pair assignments for all 637 interaction records | ### **7.1 Clip-to-Participant Mapping** The clip-to-participant mapping includes: * `clip_id` * `video_filename` * `player_1_id` * `player_2_id` * `pair_id_ordered` * `pair_id_unordered` * `movement_class` * `movement_relation` The canonical master manifest also includes participant and pair identifiers for the interaction records. ### **7.2 Internal Identity Confidence** Historical audit materials include identity-confidence and review-status fields. These describe the identity-recovery process rather than the semantics of the Ampe interaction. They remain internal quality-control information and are not included in the clean release metadata. They must not be treated as supervised prediction targets. ## **8\. Exact-Duplicate Provenance** Canonical metadata include information about exact duplicate RGB/audio media. Relevant attributes include: * Media SHA-256 * Exact-duplicate indicator * Duplicate-group identifier * Duplicate-group size The V1 audit identified: | Metric | Count | | :---- | :---- | | Exact-duplicate groups | 78 | | Records in duplicate groups | 158 | | Redundant records beyond one representative per unique hash | 80 | No exact-duplicate group contained conflicting verified movement classes. Duplicate membership is a provenance and benchmark-splitting attribute, not a semantic activity label. **Complete-file RGB/audio SHA-256 is the canonical exact-duplicate grouping unit.** Pose-file equality must not be used as a substitute because independently generated pose files for byte-identical videos may differ numerically. ## **9\. Technical Pose-Quality Annotations** Ampe-DB V1 includes technical information describing the quality and coverage of canonical normalized pose representations. Relevant quality indicators include: * Pose availability * Structural validation status * Frame coverage * Row coverage * Frame-player-group coverage * Missing frame-player groups * Partial landmark groups * Landmark visibility summaries These are technical quality-control measurements, not semantic ground-truth labels. The original canonical pose audit found 93 clips with at least one missing frame-player group relative to the source videos. Researchers applying quality filters must report the criteria, exclusions, and resulting class and participant-pair distributions. Further information is provided in `docs/POSE_FORMAT.md`. ## **10\. Technical Audio-Quality Annotations** Technical audio metadata describe recorded stream characteristics. These may include: * Audio-stream availability * Audio decoding status * Codec * Sample rate * Channel count * Channel layout * Audio/video duration difference * Signal level * Digital-silence flag * Low-level review flag These fields describe technical audio properties. They do not establish ground truth for acoustic events such as speech, clapping, footsteps, landing sounds, or rhythm. ## **11\. Pose-Derived Skeleton Outputs** The intended release includes three skeleton collections: * `skeleton/csv/` * `skeleton/json/` * `skeleton/videos/` These contain derived two-person keypoint data and rendered skeleton animations. Skeleton representations originate from the RGB pose-extraction process and, for selected clips, subsequent quality-control corrections. They are not independent human annotations and must not be treated as a separate semantic ground-truth modality. The original canonical normalized pose files in `pose/` remain the reference representation for the original V1 pose benchmark. ### **Subsequent Skeleton Corrections** Some skeleton representations were reviewed and corrected after the original processing and audit. These corrections addressed problems such as limb positioning, tracking inconsistencies, and disagreement between rendered skeleton movements and the corresponding RGB recordings. A corrected skeleton representation does not automatically establish independently observed anatomical ground truth. Likewise, correction of a skeleton CSV, JSON, or animation does not imply that the original canonical normalized pose CSV was modified. The final publication collection should retain one selected skeleton representation per canonical clip identifier and document applicable corrections and validation status. ## **12\. Historical Automatic Movement Annotations** The historical research workspace contains automatically generated movement annotations. These annotations were compared with the verified movement classes during the V1 audit. Overall agreement was approximately: **23.86%** The historical automatic movement labels are therefore not accepted as canonical V1 semantic ground truth. They should not be used as authoritative classification labels. A separate methodological study may examine the historical annotation algorithm, but it must clearly distinguish those automatic outputs from the verified movement labels. ## **13\. Historical Frame-Level Labels** Historical annotation files contain frame-level outputs generated using an automatic ankle-comparison procedure. The audit identified: **35,004 historical frame-level rows.** These values represent algorithmic estimates. They are not human-verified frame-level semantic annotations and must not be presented as canonical temporal ground truth. ## **14\. Historical Landing-Frame Predictions** Historical annotation files also include landing-frame predictions. The V1 audit found: * 402 clips with a non-null historical landing-frame value * Every non-null value was frame 2 This distribution is degenerate and does not provide credible landing-event ground truth. The historical landing-frame annotations were therefore rejected. They must not be used as verified landing labels in V1 benchmark experiments. ## **15\. Temporal Annotation Status** Canonical V1 does not include verified ground truth for: * Landing frame * Take-off frame * Movement-decision frame * Temporal lag * Synchronization timing * Foot-contact timing * Clap timing Any future event-level temporal benchmark requires a validated annotation layer with clearly defined events, annotation procedures, and quality checks. Where appropriate, manual temporal annotation should first be performed at the unique RGB/audio-content level. Annotations may then be propagated to byte-identical canonical records, provided the event timing and associated provenance remain applicable. ## **16\. Synchronization and Temporal-Lag Labels** Historical project descriptions referenced synchronization and temporal-lag annotations. The V1 audit did not identify independently verified annotations suitable for canonical ground truth for either task. Accordingly: * Synchronization classification is not a currently supported canonical labeled task. * Temporal-lag estimation does not have verified V1 target values. * Historical millisecond-level lag claims should not be carried into the publication unless independently validated. The existence of synchronized video and embedded audio is not equivalent to having human-verified synchronization-event labels. ## **17\. Winner and Round-Outcome Labels** Some benchmark artifacts include the field: `winner_participant_id` This variable is an auxiliary rule-derived outcome. It is not an independently observed or manually verified winner label. It is not part of the canonical file: `annotations/ampe_db_v1_labels.csv` Researchers must not describe it as primary V1 ground truth. Outcome interpretation also depends on rule and role semantics, including movement-token ordering that has not been independently mapped to canonical participant identity ordering. Claims about actual gameplay winners therefore require separate validation. ## **18\. Recording-Session Provenance** Canonical V1 does not contain a verified `session_id`. Historical file timestamps were investigated but rejected as a reliable proxy for capture-session provenance. Researchers must not infer recording sessions from file modification dates and present those inferences as observed session ground truth. Session-held-out benchmarks are not supported by the existing canonical provenance. ## **19\. Full-Game Recordings** The wider research workspace contains 27 longer full-game recordings. However, a verified mapping between these recordings and the 637 canonical segmented clips has not been established. Full-game identity must not be used as a verified session or clip-grouping annotation. The longer recordings are outside the canonical V1 release. ## **20\. Annotation Verification Categories** Ampe-DB distinguishes five annotation-status categories. | Category | Definition | Examples | | :---- | :---- | :---- | | **Verified** | Directly supported by the canonical audit and accepted as V1 ground truth | Movement class, pseudonymous participant identity, participant-pair membership | | **Derived** | Deterministically calculated from verified data or defined rules | Binary movement relation, auxiliary rule-derived outcome | | **Technical QC** | Describes file or representation quality rather than interaction semantics | Visibility, frame coverage, duplicate membership, audio decoding | | **Historical / Experimental** | Earlier algorithmic products not accepted as canonical ground truth | Automatic ankle-comparison labels, historical landing-frame predictions | | **Unavailable / Unverified** | Not supported by the current evidence | Verified landing time, temporal lag, recording session, independently observed winner | These categories must remain distinct in the dataset documentation and associated research publications. ## **21\. Official Supervised Benchmark Targets** Canonical V1 supports two principal classification tasks. ### **Four-Class Movement Classification** **Target:** `movement_class` Classes: * `left_left` * `left_right` * `right_left` * `right_right` ### **Binary Movement-Relation Classification** **Target:** `movement_relation` Classes: * `same_leg_configuration` * `different_leg_configuration` The binary classification label is derived from the four-class target. It should not be characterized as an additional independently observed behavior. ### **Benchmark Warning** Because participant-pair membership and movement-class distribution are strongly associated, models may exploit participant-specific or pair-specific characteristics under poorly designed evaluation protocols. Official benchmarks must therefore preserve exact duplicate grouping and report relevant participant and pair overlap. Pair-held-out evaluation is the primary V1 generalization benchmark. Participant-held-out evaluation is a diagnostic protocol restricted to eligible participants with complete target-class coverage. See `docs/BENCHMARK_PROTOCOL.md` for the published assignments and reporting requirements. ## **22\. Fixed-Prefix Experiments** Researchers may study classification using only an initial portion of each interaction clip. For example, a model may receive the first 25%, 50%, or 75% of a clip. These tasks may be described as: * Fixed-prefix movement classification * Partial-observation movement classification * Early clip-prefix classification They must not be described as pre-landing or pre-decision prediction without verified event-timing annotations. ## **23\. Adding Future Annotation Layers** Any future V1.x annotation layer should document: 1. Annotation name 2. Operational definition 3. Allowed values 4. Annotation unit 5. Annotator instructions 6. Exact-duplicate handling 7. Missing-value rules 8. Confidence or uncertainty representation 9. Quality-control procedure 10. Inter-annotator agreement, where applicable 11. Version identifier 12. Annotation provenance New annotations should not silently overwrite rejected historical annotations. Where annotation changes affect benchmark comparability, an appropriate dataset version change should be documented. ## **24\. Summary of Canonical V1 Ground Truth** Ampe-DB V1 supports four principal verified or reproducibly derived semantic structures: 1. Four-class dyadic movement configuration 2. Binary same-versus-different movement relation 3. Pseudonymous participant identity 4. Participant-pair membership Of these, the binary movement relation is deterministic rather than independently annotated. Landing timing, synchronization events, temporal lag, recording-session provenance, and independently observed winner or round outcome remain outside the verified canonical V1 ground truth. The dataset's original annotations and its subsequently corrected skeleton representations must be distinguished when reporting scientific methods or results. --- **Ampe-DB V1 — Annotation Guide** **Documentation Revision:** October 2026 **Research Authors:** Ruth Senior Biney and Isaac Kojo Yedonu Aboah