FishingROV — YOLO26x L/R 1280 (augmented) — King scallop teacher
Zoo ID: det-scallop_yolo26x_lr_1280_aug · canonical weights: best.pt (training epoch 50)
High-capacity teacher detector for King scallops, trained on left/right split
panels of 1080p survey frames upscaled to 1280 px.
FishingROV mirrors the same detector → crop → classifier pattern on two tiers
with different models. On the GPU server (RTX 3090) this teacher generates
regions of interest and feeds the cropped detections to a SwinV2 classifier.
The on-device Aura tier runs the lighter scout detector with a
MobileNetV2 classifier. This model is the 3090-side detector. The full
pipeline is still to be validated.
Metrics (honest, station-disjoint held-out)
Re-validated with model.val(imgsz=1280, conf=0.001, iou=0.6) on the public
Zenodo Test files stations — locations never seen during training.
| Metric | Value |
|---|
| mAP50 | 0.705 |
| mAP50-95 | 0.443 |
| Precision | 0.737 |
| Recall | 0.637 |
| Peak single-epoch mAP50 | 0.712 |
On data integrity. Validation panels are the public Zenodo Test files stations (station-disjoint from training) and are byte-identical to the non-augmented teacher's val set — only the training set was augmented. The reported numbers are therefore honest held-out metrics, not an inflated random-frame split.
Model details
| |
|---|
| Architecture | YOLO26x |
| Input size | 1280 px (left/right split panels) |
| Classes | 1 (scallop) |
| Train panels | 17241 (augmented) |
| Val panels | 1376 |
| Source dataset | DS-LR1280-v1-aug |
Best honest L/R teacher in the FishingROV zoo. Augmentation added ~+0.05 mAP50 over the non-augmented baseline (scallop_yolo26x_lr_1280, mAP50 0.657) on the same held-out stations.
SwinV2 classifier metrics (same-crop eval)
The 3090-tier classifier paired with this detector is SwinV2-B (256).
It was trained on DS-CLS224 (classifier_data) and evaluated on its
station-disjoint val split derived from Zenodo Test files (no random
frame mixing). Crops are square, centered on human boxes, padded if needed,
then resized to 224px; negatives are sampled away from GT boxes.
| 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
- The 3090-side detector: it generates regions of interest and feeds the
cropped detections to a SwinV2 classifier. The same detector → classifier
pattern is mirrored on the on-device Aura tier with a lighter scout detector
and a MobileNetV2 classifier (different models).
- Also usable as an offline pseudo-labelling / auto-annotation teacher to
bootstrap training data. Not a final stock-assessment instrument.
- The full pipeline is still to be validated.
- Trained only on the public St Andrews survey distribution; performance on other
gear, lighting, or substrate is unverified.
- Partially buried and king-scallop instances remain the hardest cases.
Files
best.pt — canonical weights (fitness-best epoch 50).last.pt — final-epoch weights.results.csv, results.png, curves — training history and PR/F1 curves.
Attribution & License
This model is a derivative work based on the University of St Andrews King Scallop dataset.
In accordance with the original dataset's terms, this derivative work is released under the Creative Commons Attribution 4.0 International (CC-BY 4.0) license. You are free to share and adapt this material, provided you give appropriate credit to the original authors and indicate if changes were made.
