Genuine fine-tuning in Kaggle on 256 original TilapiaVision training images and 32 original RAS training frames. Mask supervision uses only genuine RAS polygons. One fish class, not a species classifier.
Primary validation: video GX020206 (16 frames). Primary test: independent video GX020019 (16 frames). Training: GX010206 and GX030013 plus TilapiaVision train subset. No augmented sibling leakage. Test never used for checkpoint selection.
{
"protocol_id": "44380bf70d3eb10233a60dca9dcbd277804742bed6140accdabe7ab5bce04669",
"images": 16,
"map50": 0.2886014594906618,
"map50_95": 0.15082475752512942,
"precision": 0.09545454545454546,
"recall": 0.525,
"count_mae": 11.625,
"mean_iou": null,
"mean_dice": null,
"latency_median_s": 0.03987272249992202,
"vram_peak_bytes": 194312192,
"checkpoint_bytes": 80498392
}
Same Tesla T4/float32, three warmups, common max640 (RT-DETR square640), confidence0.3, bboxIoU0.5; end-to-end synchronized image timing excludes network/loading/queue. Full protocol and results in benchmark.json. Single held-out video: provisional, no farm-wide robustness claims. No trout metrics.
TilapiaVisionDataset, de Oliveira et al. (2026), CC BY4.0; Tilapia-RAS, Aragón-Banderas et al. (2025), CC BY4.0.
Derived from PekingU/rtdetr_r18vd, Apache-2.0. Original COCO classifier replaced; fish head trained, never relabeled from another COCO category. Detection only, masks unsupported.