Finetuned computer-vision models
A running showcase of the detection, segmentation, and tracking models I finetune. Each one is shown working on real footage, with the recipe it was built from and a single command to run it on your own video. No sign-up, no upload: just the results and the code.
Finetuned to flag road potholes from dashcam and street-level footage. On a battered rural road it boxed multiple defects per frame, holding pothole calls through puddles and gravel so a drive turns into a geotagged map of pavement damage.
Tuned for early domestic fire: cooking, electrical, cigarettes. A temporal state machine escalates clear, smoke present, early warning (growing smoke), fire confirmed, so it alerts on persistent, growing smoke instead of a single noisy frame.
Segments individual PV arrays from drone and nadir aerial imagery with tight masks on the panels themselves. On a Wikimedia solar-farm test it locked panels at conf 0.90 to 0.97 on 6 of 8 shots, missing only very distant obliques.
Class-agnostic tracking where pretrained detectors have no labels: a neutrophil chasing a bacterium. The pattern is human-seed-then-propagate. A few clicks lock the right object, then SAM2 propagates the rest autonomously. No training required.
No models in this category yet.