imageomics/drone-maneuver-clips

Dataset Card for Drone Maneuver Test-Clip Library

A benchmark of 41 six-second drone clips (180 frames @ 30 fps) cut from 20 KABR videos across 17 survey sessions at the Mpala Research Centre, Kenya. Each clip is indexed by which autonomous-flight maneuver it can test (launch / follow / behavior-adaptive / SoI-aware) and ships with a per-frame-per-track label table. The library is intended for evaluating drone navigation policies against real wildlife footage.

Dataset Details

This dataset repackages the KABR aerial behavior dataset into short, mix-and-match clips that each exercise a specific drone maneuver, so navigation policies can be benchmarked per-maneuver rather than only end-to-end. Labels are aligned per frame per tracked individual.

Supported Tasks and Leaderboards

Object detection / tracking, behaviour recognition, pose (viewpoint) estimation, and maneuver-conditioned navigation-policy evaluation (the primary intended use).

Dataset Structure

kabr_clips/
    catalog/
        video_index.csv      # per source video: frame summary + joined FAIR² metadata
        clip_index.csv       # one row per clip (the master index)
        coverage_report.md   # species x habitat x bbox-size x maneuver coverage
        pose_audit.csv       # per-video GT-pose assignment audit
    clips/
        <clip_id>/
            clip.mp4             # 6 s, 180 frames @ 30 fps
            labels.csv           # one row per frame per track
            maneuver_labels.csv  # per-frame ground-truth drone action, per maneuver
    README.md

The source video for each clip is referenced by its fair2_video_eventID, which resolves in the imageomics/kabr-full-release FAIR² dataset (data/video_events.csv); the session it belongs to is fair2_session_eventID (data/session_events.csv). The catalog carries no absolute filesystem paths.

clip_id = <date>-<video>_<start_frame> (e.g. 18_01_2023_session_7-DJI_0070_000360).

Data Fields

clips//labels.csv (one row per frame per track):

• clip_id, video_id, session_id: identifiers; session_id is the KABR/FAIR² session event.
• frame_global: absolute frame in the source video; frame_local: 0-based within the clip; time_s.
• track_id: KABR mini-scene id, persistent within the source video.
• species, behaviour: KABR expert ground-truth labels; vigilant: behaviour in {Head Up, Running, Trotting}.
• pose, pose_provenance, pose_match_score: 8-class viewpoint where ground-truth exists (see Annotations); else empty.
• individual_id: empty — global re-identification is future work (no ground truth exists).
• xtl,ytl,xbr,ybr, x_c,y_c,w,h, bbox_area_frac, bbox_size_class (far/medium/close, relative to this survey).
• occluded, outside: KABR annotation flags; latitude, longitude, altitude: drone telemetry; date_time.

catalog/clip_index.csv (one row per clip): identifiers, frame range, species_set, habitat, habitat_notes, habitat_provenance, herd_size, behaviours_present, has_vigilance, bbox_size_classes, pose_set, suitable_maneuvers, and the FAIR² event IDs (fair2_video_eventID / fair2_session_eventID) that point back to the source video in imageomics/kabr-full-release.

habitat is a structural class — open, closed, mixed, or unknown — so it can be filtered and grouped consistently. It is derived from the original free-text field metadata (Bitterlich relascope scores — a basal-area count, higher = denser woody canopy — plus field remarks), which is preserved verbatim in habitat_notes. The binning is: Bitterlich 1–2 and explicit "open" → open; Bitterlich 3–4, "mix", scattered/some bushes → mixed; Bitterlich ≥5 (e.g. 10) and "bushy" → closed; bare "near watering hole" / blank → unknown.

habitat/habitat_notes and herd_size are joined from the FAIR² session_events. habitat_provenance records where each original habitat value came from: session_event (the structured habitat field), eventRemarks (best-effort recovery from the free-text session remarks where the structured field was blank), or empty (no habitat information available for that session).

clips//maneuver_labels.csv (one row per frame per maneuver) — the ground-truth drone action produced by replaying the formal maneuver decision tree over labels.csv:

• maneuver: one of approach / track / behavior_adaptive / soi_aware.
• action_set_raw, action_set_smoothed: the per-frame action(s) from the 9-action space (up, down, forward, back, left, right, yaw-left, yaw-right, hover); smoothed is the published label after a 3 s rolling average that suppresses jitter.
• triggering_branch: which decision-tree branch fired (for auditability).
• S_t, pct_vigilant, centroid_x, centroid_y, mean_px, n_tracks: the frame features the decision used, so any label is reproducible and inspectable.

Evaluation harness (maneuver decision tree)

The accompanying replay harness (clip_library/maneuver_labels.py, spec in maneuver_decision_tree.md) executes a small, inspectable policy specification deterministically over each clip, emitting the action an expert-calibrated controller would take per frame. A learned navigation policy can then be scored, per maneuver, against this reference. Every threshold (vigilance theta_S, desired bbox pixels, SoI pose, smoothing window) is user-tunable; tuned runs write a maneuver_labels.custom.csv sidecar and never overwrite the released labels. The harness is CPU-only and runs on the released clips without the raw KABR archive.

Composition

• Clips: 41 across 20 videos / 17 sessions.
• Maneuver coverage:
  • follow: 39 clips
  • soi_aware: 36 clips
  • behavior_adaptive: 23 clips
  • launch: 11 clips
• Species (clip-level membership; Zebra is the coarse KABR label, see note below):
  • Zebra: 18 clips
  • Grevys Zebra: 17 clips
  • Giraffe: 15 clips
  • Plains Zebra: 6 clips
• Bbox size class (relative to this survey's range):
  • medium: 32 clips
  • far: 24 clips
  • close: 20 clips
• Habitat structural class (original free text in habitat_notes):
  • closed: 15 clips
  • mixed: 13 clips
  • open: 11 clips
  • unknown: 2 clips
• Clips with ground-truth pose: 26 (5 pose-annotated videos).

Dataset Creation

Curation Rationale

Prior drone-ecology evaluation is end-to-end and offline. This library isolates the maneuvers an autonomous drone must perform (approaching, following a herd, responding to disturbance, capturing a desired viewpoint) into short replayable clips with aligned ground truth, so navigation policies can be benchmarked per maneuver under realistic perception conditions.

Source Data

KABR full release (Mpala Research Centre drone surveys, January 2023; DJI aircraft, 20–50 m altitude). Clips are cut from the original videos; labels are joined from the KABR per-frame occurrence records. See KABR: https://doi.org/10.48550/arXiv.2510.02030

Annotations — label provenance (read this)

Labels come from three distinct sources, stated explicitly:

1. Bounding box, species, behaviour — KABR expert ground truth. Frame-by-frame CVAT annotations from the KABR project; carried over verbatim (duplicate merge-rows collapsed to one row per track per frame).
2. Pose (8-class viewpoint) — ground-truth cross-reference, sparse. Sourced from the manually labeled imageomics/KABR-poses crops (5 videos: DJI_0002, DJI_0006, DJI_0070, DJI_0145, DJI_0208). We deliberately do not run the DINOv2 pose classifier here: it was trained on these same KABR crops, so applying it would be inference on training data. Because a crop's filename identifies its frame but not its track, each labeled crop is matched to a KABR track by visual disambiguation among the candidate boxes at that frame; matches below a similarity threshold are flagged pose_provenance = "gt-ambiguous". Of 227 crop assignments, 125 were flagged ambiguous (see catalog/pose_audit.csv). Pose coverage is sparse by design.
3. Individual ID — absent (future work). No global re-identification ground truth exists; only the within-video track_id is provided. Cross-video re-ID (e.g. MegaDescriptor) is left for future work.

Who are the annotators?

KABR project annotators (bounding boxes, behaviour); the imageomics/KABR-poses curator (pose). Maneuver-suitability tags are assigned programmatically by the selection pipeline.

Personal and Sensitive Information

Wildlife only; no personal data. Species include Grevy's zebra (endangered) — coordinates are at survey-session granularity.

Considerations for Using the Data

Bias, Risks, and Limitations

• Pose is sparse (5 videos) and best-effort track-matched; trust the pose_provenance/pose_match_score columns.
• Species skew toward Grevy's zebra; single site (Mpala); small bbox sizes throughout (20–50 m altitude), so "close/medium/far" are relative to this survey, not absolute scale.
• Labels are KABR ground truth, not model-generated — appropriate for ground-truth evaluation, not for characterizing perception error.
• Species/behaviour strings are normalized (Walking→Walk, Grevy→Grevys Zebra). The generic Zebra label is retained from KABR and means undifferentiated zebra — Plains and/or Grevy's: on those sessions the annotators did not split the two subspecies, so a Zebra clip may contain Plains zebra, Grevy's zebra, or both. This is a labeling-granularity difference, not noise; the explicit Plains Zebra / Grevys Zebra labels are used only where KABR distinguished them.
• Because GT pose is sparse, the SoI-aware maneuver labels are mostly hover (no-pose) on this release; the maneuver is fully exercised only with dense (model-generated) pose.

Recommendations

Filter on pose_provenance == "gt" for confident pose; use suitable_maneuvers to select clips per maneuver; cross-reference session_id to the KABR/FAIR² metadata for habitat and herd context.

Licensing Information

This dataset (the compilation) is released under CC-BY-4.0. Please also cite the original KABR and KABR-poses datasets.

Citation

BibTeX:

Data

@misc{drone_maneuver_clips,
  author = {Kline, Jenna and others},
  title = {Drone Maneuver Test-Clip Library},
  year = {2026},
  url = {https://huggingface.co/datasets/imageomics/drone-maneuver-clips},
  publisher = {Hugging Face}
}

Please also cite the original data source(s):

@misc{kabr2023,
  title  = {KABR: In-Situ Dataset for Kenyan Animal Behavior Recognition from Drone Videos},
  url    = {https://doi.org/10.48550/arXiv.2510.02030}
}

Acknowledgements

This work was supported by the Imageomics Institute, which is funded by the US National Science Foundation's Harnessing the Data Revolution (HDR) program under Award #2118240 (Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

Dataset Card Authors

Generated by the clip_library pipeline; curated by Kline et al.

Dataset Card Contact

See the project repository.