WhaleDrone: Los Cabos Humpback Whale UAV Dataset

Prepared following the FAIR2Drone dataset standard.

Summary

WhaleDrone is a UAV-collected dataset of humpback whale (Megaptera novaeangliae) competitive groups filmed off Los Cabos, Baja California Sur, Mexico, in January 2026. It comprises 26 flights (~102 RGB video clips), per-clip telemetry, flight-level Airdata logs, and a sparse keyframe annotation set covering 17 individual whales, a reference vessel, and dolphins across 5 fully annotated flights.

WhaleDrone is released as the data companion to the WhaleTrack pipeline paper (MarineVision Workshop, ECCV 2026), however the footage, telemetry, and annotations are equally suited to other detection, tracking, or behavioral-ecology work. Code, trained detection/tracking models, and train/val/test splits used in the WhaleTrack paper are released separately — see Related resources. The CV methods paper cites both resources.

Dataset at a Glance

Location	Los Cabos, Baja California Sur, Mexico
Collection period	January 2026
Platform	DJI Mavic 3T (RGB wide + thermal, synchronized)
Flights	26
Video clips (RGB)	~102
Annotated flights	5
Annotated individuals	17 humpback whales (A001–A017), 1 reference vessel (7.3 m), dolphins
Telemetry	1 .SRT per RGB clip (GPS, altitude, gimbal pitch/yaw), per-flight Airdata CSV
Coordinate system	UTM Zone 12Q; WGS84 (geodetic)
Permit	Dirección General de Vida Silvestre (DGVS), Official Letter No. SBRA/DGVS/03662/25
Ethics approval	SDU Research Ethics Committee, Approval No. 25/66344
License	CC-BY-NC-4.0
Standard	FAIR2Drone; metadata tables follow Darwin Core (Event/Occurrence)

Repository Structure

WhaleDrone/
│
├── DATASET_CARD.md                          ← this file
├── MX_2026_platform_specs.csv               ← full sensor & platform specifications
├── camera_calibration_DJIM3T_RGBwide.json   ← K matrix, distortion coefficients, Python snippet
├── dataset_index.xlsx                       ← all 26 flights: DVM name ↔ DJI filename, annotation status, group composition
├── TECHNICAL_README.md                      ← annotation format, CSV columns, metric-computation recipes
│
├── metadata/
│   ├── DC_events.csv                        ← Darwin Core Event table (1 row per video)
│   └── DC_occurrences.csv                   ← Darwin Core Occurrence table (1 row per individual × video)
│
├── annotations/
│   └── annotations_<DVM-project-name>.csv   ← annotation file (point, vector)
│
├── telemetry/
│   ├── README_time_sync.md                  ← timestamp offsets across .SRT / Airdata / local time
│   ├── DJI_YYYYMMDDhhmmss_NNNN_V.SRT        ← one per RGB clip (per-frame GPS, altitude, gimbal)
│   └── *-Flight-Airdata.csv                 ← one per flight, full telemetry log (UTC+0)
│
└── videos/
    └── DJI_YYYYMMDDhhmmss_NNNN_V.MP4         ← RGB clips

File-by-File Description

File / folder	Content	Primary use
MX_2026_platform_specs.csv	Full drone + sensor specs: all 3 cameras (RGB wide, thermal, side-by-side), gimbal, GNSS	Methods-section reference
camera_calibration_DJIM3T_RGBwide.json	K matrix, distortion coefficients, Python undistortion snippet	Undistortion + homography
dataset_index.xlsx	All 26 flights: DVM internal name → original DJI filename, annotation status, group composition	File lookup, dataset overview
TECHNICAL_README.md	Full annotation format, CSV column definitions, metric-computation recipes (Python)	Reproducing derived metrics from raw annotations
metadata/DC_events.csv	Darwin Core Event core — one row per video clip	Interoperable event-level metadata (GBIF/OBIS-compatible)
metadata/DC_occurrences.csv	Darwin Core Occurrence extension — one row per individual × video	Interoperable occurrence-level metadata
annotations/annotations_*.csv	Sparse keyframe annotations, one CSV per flight, raw DVM export: point rows (position only) and vector rows (tail-to-rostrum length + heading) per individual per frame	Ground-truth validation, behavioral coding
telemetry/README_time_sync.md	Documented offsets between drone clock, Airdata CSV, and local time	Cross-file time alignment
telemetry/*.SRT	Per-frame GPS position, altitude, gimbal pitch/yaw — one file per RGB clip	Georeferencing
telemetry/*-Flight-Airdata.csv	Full per-flight telemetry log (UTC+0)	DVM sync validation; fallback telemetry source
videos/*.MP4	RGB video only, ~102 clips across 26 flights	Detection + tracking input

Coordinate & Time Conventions

• Spatial: UTM Zone 12Q for all projected coordinates in annotation and Darwin Core tables; WGS84 for raw GPS in .SRT/Airdata.
• Temporal: the drone's internal clock is CET / UTC+1 (not CEST); Airdata CSV timestamps are UTC+0; local time at Los Cabos is UTC−7. Offsets were validated empirically against a simultaneous iPhone recording. See telemetry/README_time_sync.md for the full per-flight offset table and worked conversion example.
• Scale/velocity validation: a reference vessel of known length (7.3 m), filmed at nadir, is included in the annotated flights and used to validate reprojected distance/speed against ground truth.

Known Data Quality Notes

• DroneVideoMeasure (DVM) renames video files internally on import. dataset_index.xlsx carries the authoritative DVM-name ↔ original-DJI-filename mapping; always join through this table rather than assuming filename correspondence.
• Some annotation rows are vector rows (tail-to-rostrum, length populated) and some are point rows (body midsection, length = NaN) — see TECHNICAL_README.md for the full column semantics before computing derived metrics.

Darwin Core Compliance

This dataset follows the Darwin Core standard for biodiversity data exchange.

Event records (metadata/all_events.csv): one row per video file. Fields include eventID, eventDate, eventTime (UTC), decimalLatitude, decimalLongitude (rounded), samplingProtocol, samplingEffort, and platform telemetry fields. Occurrence records (metadata/occurrences.csv): one row per individual × video. Fields include occurrenceID, scientificName (full taxonomy to species level), individualID, lifeStage, individualCount, basisOfRecord, and identificationRemarks. Coordinates are rounded to 2 decimal places (~1 km) to protect exact animal locations.

Use Case

WhaleDrone was collected and annotated as the foundational dataset for a one-day hackathon organized with the IDEFIX team (IRISA, Université Bretagne Sud, Vannes, FR). The hackathon produced WhaleTrack, a drone-based multi-whale tracking pipeline that detects whales in video, tracks them across frames, and reprojects their positions into real-world GPS coordinates using drone telemetry alone (GPS, altitude, gimbal angle), without visual ground control points. The pipeline also computes a per-frame field-of-view coverage polygon, distinguishing true absence of an individual from an area simply not observed by the drone at that moment.

WhaleTrack was presented at the 2nd MarineVision Workshop, ECCV 2026, in Malmö, Sweden - 08th Sep, 2026.

Hackathon participants: Pierre Adorni, Matthieu Le Lain, Thomas Fillon, Louis François–Downey, Kshitij Raj Sharma, Shivam Pande, Manuel Nkegoum, Sergio Suzerain Osson, Emanuel Goulart Farias, Frédéric Raimbault, Charlotte Pelletier, Marc Chaumont, and Sébastien Lefèvre. Chair: Lucie Laporte-Devylder

ML-ready dataset: [link once published] Code / models (GitHub): [link once published]

Related Resources

• Models / train-val-test splits: https://huggingface.co/datasets/pierreadorni/WhaleDrone-YOLO — WhaleTrack detection, tracking, and georeferencing pipeline.
• Code: https://github.com/pierreadorni/WhaleTrack
• Paper: https://openreview.net/forum?id=EWYJA9b6oJ

Authors & Credit

Name	Role (CRediT)	Affiliation
Lucie Laporte-Devylder	Conceptualization, Investigation (field data collection), Data curation, Methodology, Annotation, Writing	University of Southern Denmark / WildDrone doctoral network
Esther Jimenez	Investigation (field data collection), Resources (permit, vessel, local logistics), Writing – review	Universidad Autónoma de Baja California Sur (UABCS), La Paz, Mexico / MMAPE
Hiram Rosales Nanduca	Investigation (field data collection), Resources (permit, vessel, local logistics), Writing – review	Universidad Autónoma de Baja California Sur (UABCS), La Paz, Mexico / MMAPE

Citation

@dataset{lucie_laporte-devylder_2026,
    author       = { Lucie Laporte-Devylder },
    title        = { WhaleDrone: Los Cabos Humpback Whale UAV Dataset },
    year         = 2026,
    url          = { https://huggingface.co/datasets/LucieLprt-Dvldr/WhaleDrone },
    doi          = { 10.57967/hf/9951 },
    publisher    = { Hugging Face }
}

Funding

This work was supported by the WildDrone MSCA Doctoral Network funded by EU Horizon Europe under grant agreement no. 101071224.

Contact

Lucie Laporte-Devylder — [lucie@biology.sdu.dk] — University of Southern Denmark / WildDrone