Ganaraska River: Live Model Outputs Experimental

DINOv3, the vision backbone inside the model, maps each fish into a high-dimensional “fingerprint” space where same-species fish land together. Above are 1,000 real fish from the validation set, coloured by species — that space squeezed down to 3 so you can spin it around. The clean clusters are why the model can tell them apart; notice Chinook and Coho, the trickiest pair, sitting closest.

The model is DEIMv2-X, a real-time object detector built on a DINOv3 vision backbone that was pre-trained on 1.7 billion images, so it already understands shape, colour and texture. We fine-tuned it on thousands of fish frames that technicians had labelled by species, across both rivers and several seasons, so that one network both finds the fish and names it.

• Detector + classifier — one DEIMv2-X network (DINOv3 ViT-S+/16 backbone) boxes and names each fish in a single pass.
• Tracker — ByteTrack links a fish across frames to count it once and read its direction.
• Accuracy — about 96% species agreement on 366 held-out fish it never trained on.

A research tool, not an official count. The “Ganaraska River” caption in the clips is burned in by the camera itself.