shivam11/grand-bassin-kanwar

Grand Bassin Kanwar Dataset

An anonymised, auto-annotated image dataset of the Maha Shivaratri pilgrimage at Grand Bassin (Ganga Talao), Mauritius, focused on pilgrims and the devotional bamboo-and-flower structures (kanwars, also spelt kawars) they carry. To our knowledge this is the first public computer-vision dataset of this event.

Images	2,054 JPEG frames, 1280×720
Annotations	98,387 bounding boxes, COCO format
Streams	21 video segments from 9 camera sources
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

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

Each COCO image entry's file_name is the path relative to images/ (forward slashes). Frames within a stream are consecutive, sampled at 2.0 fps from the source video.

Categories

Eleven categories are declared in the COCO file; three contain instances in this release:

id	name	boxes	source
1	pilgrim	57,408	YOLO26m auto-annotation (COCO person)
5	vehicle	34,852	YOLO26m auto-annotation (all COCO vehicle classes merged)
3	kanwar	6,127	HSV colour saliency + MobileNetV3 classifier (see below)

The remaining eight (kanwar_carrier, police_officer, kanwar_decorated, crowd, vendor, devotional_items, water_container, trishul) are declared with zero instances — they are the intended taxonomy for future manual annotation and are kept so that annotation tooling and category ids remain stable across releases.

Kanwar boxes appear in 2,003 of 2,054 images. Per-image annotation counts range 1–129 (median 47).

Camera sources

Stream prefix	View	Frames	Description
stream_gbassinpetrin.* (13 segments)	Ground / roadside CCTV	2,000	Le Petrin road toward Grand Bassin; the primary kanwar procession route
vid_from_q1, q5–q9	Ground / roadside	32	Auxiliary road cameras (auto-classified)
vid_from_q2, q4	Aerial / overhead	22	Overhead parking-area views (auto-classified)

Camera view labels for the vid_from_* streams were assigned by a Gaussian-Mixture-Model camera-type classifier (99.46% accuracy against ground-truth stream names); details in the companion paper.

Annotation process — read before benchmarking

All annotations are machine-generated. None are human-verified.

• pilgrim / vehicle: Ultralytics YOLO26m, imgsz=1280, confidence 0.15, COCO classes remapped (person→pilgrim; bicycle, car, motorcycle, bus, truck→vehicle).
• kanwar: two-stage pipeline — HSV colour-saliency proposal (orange / saffron / red regions) followed by a MobileNetV3-Small binary classifier (98.8% validation accuracy on a small single-camera split, acceptance threshold 0.60). These boxes are best treated as high-recall candidates, not ground truth.

Suitable uses: pretraining, semi-supervised and weakly-supervised learning, label-noise research, domain adaptation, crowd/event analysis. If you need gold labels, budget for re-annotation — the declared taxonomy above is designed for it.

Privacy and anonymisation

All frames were processed on 2026-07-12 with an OpenCV-based anonymisation pipeline (anonymise_kanwar.py):

• Road frames: YuNet face detection (conf 0.50) + Haar plate & frontal-face cascades, plus annotation-driven blurring of the top 50% of every person-type box and the rear-plate zone (bottom-centre 15%) of every vehicle box.
• Aerial frames: YuNet at conf 0.35 + full person-bbox blurring; vehicle plates are not legible at this altitude and are left intact.
• Blurring is pixelate + Gaussian; box geometry is unchanged, so annotations remain valid.
• Totals: 3,538 face regions and 33,391 plate regions redacted across 2,054 frames, zero errors. A per-run audit summary is retained by the maintainers.

Kanwar structures themselves are never blurred.

Detection-based anonymisation is not perfect; a face or plate missed by both the detectors and the auto-annotations may remain legible. 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.

Ethical considerations

The imagery documents a religious observance. It is published for research on crowd safety, traffic management, and cultural-event computer vision. Users should treat the subject matter respectfully: no mockery or misrepresentation of the depicted practices, and no attempts to profile participants.

Limitations

• Single event, single day, one primary viewpoint: 97% of frames come from one roadside camera. Models trained here will not generalise to other scenes without adaptation.
• Temporal correlation: frames are 0.5 s apart within a segment. Consecutive frames are near-duplicates — split train/val/test by segment, never by random frame, or results will be inflated.
• Auto-annotation noise: expect missed detections (especially small or occluded pilgrims), duplicate boxes, and colour-triggered kanwar false positives (e.g. orange clothing, flags).
• Weather/lighting variety is limited to daytime conditions on the event days.

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.21321320. A dataset descriptor paper is in preparation; until it is published, cite:

@misc{grandbassin_kanwar_2026,
  title     = {Grand Bassin Kanwar Dataset: Anonymised Imagery of the
               Maha Shivaratri Pilgrimage, Mauritius},
  author    = {Guness, Shivanand Prabhoolall},
  year      = {2026},
  doi       = {10.5281/zenodo.21321320},
  publisher = {Zenodo},
  note      = {Version 1.0. Descriptor paper in preparation.}
}

Maintainer

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

Changelog

• v1.0 (2026-07-12) — initial anonymised release candidate: 2,054 frames, 98,387 auto-annotations, faces and plates redacted.