FishingROV/classifier_swinv2b_256

FishingROV - SwinV2-B 256 classifier (crops)

Zoo ID: cls-classifier_swinv2b_256 - canonical weights: best.pt

SwinV2-B classifier for 5-way scallop taxonomy on square crops. Used in the 3090-side detector -> crop -> classifier pipeline (paired with the YOLO teacher for ROI generation). This is a classifier only; it does not perform detection.

Dataset provenance (station-disjoint)

This model is trained on DS-CLS224 (classifier_data) generated from the public Zenodo dataset:

• Train split: Zenodo Training files stations
• Val split: Zenodo Test files stations (station-disjoint)

Crops are square, centered on human boxes, padded if needed, resized to 224px. Negatives are sampled away from GT boxes. No augmentation.

Metrics (same-crop validation)

All metrics below are from classifier_data/val (station-disjoint Test files).

Metric	Value
Macro precision	0.700
Macro recall	0.654
Macro F1	0.661
Accuracy	0.966

Per-class metrics (from class_eval_best.json):

Class	Precision	Recall	F1	Support
dead	0.464	0.642	0.539	81
king	0.391	0.237	0.295	76
not_a_scallop	0.991	0.996	0.993	5781
queen	0.818	0.899	0.857	296
recessed	0.837	0.497	0.623	145

Intended use & limitations

• Intended for crop classification, not end-to-end detection.
• Trained only on the St Andrews survey distribution; generalization is unknown.
• King class remains the weakest; expect confusion in borderline cases.

Files

• best.pt - best checkpoint (SwinV2-B 256)
• last.pt - final checkpoint
• class_eval_best.json - macro + per-class metrics
• classes.json, history.json

Attribution & License

This model is a derivative work based on the University of St Andrews King Scallop dataset.

• Original DOI: https://doi.org/10.5281/zenodo.10156830

Released under CC-BY 4.0 with attribution to the original authors.