A box edge as a distribution
Modern detection heads do not regress a distance; they emit a distribution over bins and take its expected value. Paint the histogram and the decoded edge follows, which makes the failure mode of that choice visible.
Try this
Load the bimodal preset. The expectation lands in the gap between the two modes, where almost no bin has any mass at all.
Drag across the bars to reshape the distribution.
Where this lab comes from
YOLO-NAS Strikes Back: Still the Fastest Detector on Hardware You Already OwnQuantized to INT8, YOLO-NAS-S runs at 270 FPS on a laptop CPU and 1146 FPS on a mid-range laptop GPU, and it drops straight into Frigate. The model aged well; its tooling did not. Here is a clean reimplementation, measured across CPU, Intel iGPU and NVIDIA dGPU, and checked bit-for-bit against the original.
More in detectors and trackers
- Matching detections to tracks — Drag next frame's boxes and watch identities follow — or swap.
- Reading a detector's module graph — Rows are execution steps; columns are feature-map resolution.