ratapan-ngawi/canisight-visual-signs-v4-scene-corrected

CaniSight Visual Signs V4 — Scene-Corrected Pilot

Small pilot object-detection dataset used for the CaniSight semifinal prototype. It contains dog images annotated for two observable visual signs:

• dropped_jaw
• visible_hypersalivation

This dataset does not label or diagnose rabies. These signs are nonspecific and must not be interpreted as proof that an animal has or does not have rabies.

Contents

The downloadable ZIP contains a YOLO-format dataset with train/, valid/, and test/ image/label folders plus data.yaml.

Split	Images	dropped_jaw boxes	visible_hypersalivation boxes	Null / hard-negative images
Train	46	16	28	10
Validation	11	4	3	4
Test	8	4	3	3
Total	65	24	34	17

Class mapping in data.yaml:

0: dropped_jaw
1: visible_hypersalivation

What changed from V3

V4 uses the same 65-image pilot pool as V3, but corrects a known split issue. One frame from the source scene titled Watch out for rabies infected dog symptoms... was moved from train to test so that all three frames from that scene are kept together in test. The split therefore changed from 47/11/7 to 46/11/8.

Frames were grouped by identifiable source video/scene when constructing the corrected split. Because the pilot was assembled from heterogeneous sources and does not include a complete machine-readable source/scene manifest, the scene-separation claim should be treated as a documented manual correction rather than a fully automated guarantee.

The test split was used exploratorily to compare candidate YOLO11n and YOLO11s models. It is not an untouched independent final evaluation set.

Intended use and limitations

This is a small proof-of-concept dataset intended for research, reproducibility, and demonstration of the CaniSight prototype. It is not clinically validated and is not suitable for diagnosis, autonomous veterinary decisions, or claims of rabies status.

Important limitations include the small sample size, limited breed and condition coverage, multiple frames originating from some videos, uncertain capture conditions, and sensitivity to angle, blur, lighting, occlusion, mouth size, panting, wet fur, and visually ambiguous saliva.

Rights and attribution

The pilot combines manually collected frames and public-source material with heterogeneous original terms. Public availability of this repository does not override the copyright or license of an underlying source. Users are responsible for checking each original source's terms before redistribution or commercial use. The annotations and packaging are supplied for CaniSight competition reproducibility; no blanket license is asserted over third-party images.

Integrity

SHA-256 of CaniSight_Visual_Signs_V4_scene_corrected_clean.zip:

CEBFAF32381DBA21C8357C198A4E64297DCFBB3305B415EB8125CCEC9C371979

Related model

The selected V4 pilot checkpoint and evaluation artifacts are hosted at ratapan-ngawi/canisight-yolo11s-visual-signs-v4-pilot.