Buy the pack — $149vehicles in traffic.
- detr-r50-fp32 — calibrated
- F1 0.663
- same checkpoint — stock
- F1 0.597
delta +0.066 F1 — COCO val2017 vehicle subset (250 images, 971 boxes), 800 px letterbox, 1-core x86, run 2026-09-28.
Car, motorcycle, bus and truck on road scenes. We raced every Apache-2.0 architecture on the held-out set, calibrated the winner, and ship it as a pack with the numbers attached. The full table is below, and you can check our work on the real street frame below.
The run, on a real frame.
This is the winning detector, unedited output: nine vehicles found, nothing invented. The image was not uploaded anywhere — the boxes below were drawn from the actual run and shipped with the page.
The race.
F1 at IoU 0.50 on the held-out set, stock threshold 0.5 flat versus calibrated. Latency is the median of a single CPU core.
| config | precision | recall | F1 stock | F1 calibrated | latency |
|---|---|---|---|---|---|
| detr-r50-fp32 — winner | 0.732 | 0.606 | 0.597 | 0.663 | 1606 ms |
| rfdetr-n-fp32 | 0.704 | 0.609 | 0.588 | 0.653 | 325 ms |
| rfdetr-n-int8 | 0.720 | 0.595 | 0.586 | 0.652 | 224 ms |
| detr-r50-int8 | 0.697 | 0.549 | 0.550 | 0.614 | 1332 ms |
run COCO val2017 vehicle subset (250 images, 971 boxes), 800 px letterbox, 1-core x86, run 2026-09-28. the winning config's calibrated thresholds: car 0.90, motorcycle 0.65, bus 0.90, truck 0.85 (COCO class id 3), (COCO class id 4), (COCO class id 6), (COCO class id 8); NMS IoU 0.70. the values that produced the delta are in the pack.
In the pack.
- fp32 and INT8 ONNX weightsthe winning checkpoint, both precisions, exported and decode-tested
- verified decoderPython and JavaScript inference with the output contract checked against the reference implementation
- calibration reportinput resolution, per-class thresholds, NMS IoU and the measured delta against stock on COCO val2017 vehicle subset (250 images, 971 boxes)
- Apache-2.0 attribution noticesno AGPL anywhere in the tree
- 10,000 hosted API callstop-ups never expire