shivam11/grand-bassin-traffic

Grand Bassin Traffic Dataset

An anonymised, auto-annotated image dataset of vehicle and pedestrian traffic during the Maha Shivaratri pilgrimage at Grand Bassin (Ganga Talao), Mauritius, captured simultaneously from an overhead (near-nadir) parking-area camera and a roadside CCTV camera. The paired viewpoints make it suitable for studying the aerial-vs-ground domain gap in object detection; the companion dataset (Grand Bassin Kanwar Dataset) covers the pilgrimage-specific content from the same event.

Images	4,000 JPEG frames, 1280×720 (2,000 aerial + 2,000 road)
Annotations	854,646 bounding boxes, COCO format, with per-box detector and score
Streams	24 video segments from 2 cameras
Event	Maha Shivaratri, 16–17 February 2023
Location	Grand Bassin / Le Petrin area, Mauritius
Anonymisation	Faces and licence plates pixelated (see § Privacy)
Licence	CC BY-NC-SA 4.0

Dataset structure

traffic/
├── images/<stream_name>/<hash>_f<index>.jpg
├── annotations/instances_all.json          # COCO 1.0
└── DATASET_CARD.md

Each COCO image entry carries camera_view (aerial | road), video_source, source_type, and frame_index. file_name is the path relative to images/ (forward slashes). Frames within a segment are consecutive, sampled at 2.0 fps.

Camera sources

Camera	View	Segments	Frames	Description
gbassinexch1	Aerial / overhead	11	2,000	North parking area, near-nadir, ~30 m; dense parked and moving vehicles
gbassinpetrin	Ground / roadside	13	2,000	Le Petrin road toward Grand Bassin; oblique perspective, pilgrims and traffic

Categories and annotation provenance

Eight categories are declared; six contain instances (traffic_light and stop_sign are declared but empty). All annotations are machine-generated; none are human-verified. Two detectors were used, routed by camera type:

• Aerial frames — VisDrone-finetuned YOLOv8 with SAHI sliced inference (640×360 tiles, 20% overlap, confidence 0.40). These boxes carry "detector": "visdrone_sahi" and a score field. This strategy was chosen after a controlled comparison showed statistically significant improvements over a COCO-trained baseline on overhead imagery (Wilcoxon p < 10⁻⁵); see the companion paper.
• Road frames — COCO-trained YOLO26m (imgsz 1280, confidence 0.12–0.15). These boxes have no detector field.

Category	Aerial	Road	Total
car	659,158	28,580	687,738
person	42,756	56,550	99,306
truck	36,366	4,050	40,416
bus	25,261	20	25,281
motorcycle	1,025	47	1,072
bicycle	683	150	833

Per-image box counts range 24–456 (median 199) — the aerial parking scenes are extremely dense. Aerial truck counts are inflated by VisDrone's van class being mapped to truck; treat fine-grained vehicle types on aerial frames with caution.

Privacy and anonymisation

All 4,000 frames were processed on 2026-03-19 with an OpenCV-based pipeline (anonymise_v3.py), superseding an earlier annotation-only pass:

• Road frames: YuNet face detection (conf 0.50) + Haar plate & frontal-face cascades, plus annotation-driven blurring of the top 50% of every person box and the rear-plate zone (bottom-centre 15%) of every vehicle box.
• Aerial frames: YuNet at conf 0.35 + full person-bbox blurring; plates are not legible at this altitude and vehicles are left intact.
• Blurring is pixelate + Gaussian; box geometry is unchanged, so annotations remain valid. Totals: 45,978 face regions and 32,854 plate regions redacted, zero errors. A per-run audit summary is retained by the maintainers.

Detection-based anonymisation is not perfect. If you find an identifiable person or plate, contact the maintainer (below) and the frame will be re-processed or withdrawn. Do not attempt to identify individuals; do not use this dataset for face recognition, re-identification, or surveillance of individuals.

Suggested uses

Small-object detection and sliced-inference research, aerial↔ground domain adaptation, camera-type classification, parking occupancy and crowd/traffic density estimation, label-noise and semi-supervised learning. For gold-label evaluation, budget for re-annotation.

Limitations

• Two cameras, one event, daytime only: models trained here will not generalise to other scenes without adaptation.
• Temporal correlation: frames are 0.5 s apart within a segment — split train/val/test by segment, never by random frame, or metrics will be inflated by near-duplicate leakage.
• Auto-annotation noise: the aerial detector was selected for high recall on small objects; expect residual false positives and duplicate boxes. Road-frame annotations come from a COCO-trained model with known underdetection on distant objects.
• A small set of auxiliary low-frame-count streams (vid_from_q1–q9, 54 frames) from the same recording pool is not included in this release; it appears only in the internal research corpus.

Licence

Released under Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). You may share and adapt the dataset for non-commercial purposes with attribution, provided derivatives carry the same licence. The privacy and ethical-use conditions in this card apply in addition to the licence terms.

Citation

Archived at Zenodo: doi:10.5281/zenodo.21321449. A dataset descriptor paper and a companion paper on the aerial detection methodology are in preparation. Until they are published, cite:

@misc{grandbassin_traffic_2026,
  title     = {Grand Bassin Traffic Dataset: Anonymised Aerial and Roadside
               Traffic Imagery from the Maha Shivaratri Pilgrimage, Mauritius},
  author    = {Guness, Shivanand Prabhoolall},
  year      = {2026},
  doi       = {10.5281/zenodo.21321449},
  publisher = {Zenodo},
  note      = {Version 1.0. Descriptor paper in preparation.}
}

Maintainer

Shivanand Prabhoolall Guness (ORCID 0000-0002-8808-3722) — [email protected]

Changelog

• v1.0 — 4,000 frames (2,000 aerial + 2,000 road), 854,646 auto-annotations (VisDrone+SAHI aerial re-annotation, 2026-03-18); model-based anonymisation pass v3 (2026-03-19).