Buy the pack — $149Vehicles in traffic.
- detr-r50-fp32 — calibrated
- F1 0.626
- same checkpoint — stock
- F1 0.567
delta +0.059 F1 — COCO val2017 vehicle subset (250 images, 971 boxes), 800 px, 1-core x86, run 2026-09-17.
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 your own photo.
Check it on your own photo.
Drop an image below. A stock detector runs in your browser — nothing is uploaded. The calibrated run happens on our server, because the tuned parameters are the product.
drop a .jpg or .png here — browser detection ships with the next pack release. until then the numbers above are the run.
every number on this page: dataset, resolution, hardware and date attached. this run: COCO val2017 vehicle subset (250 images, 971 boxes), 800 px, 1-core x86, run 2026-09-17.
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 | 0.699 | 0.566 | 0.567 | 1091 ms | |
| detr-r50-int8 | 0.640 | 0.505 | 0.516 | 960 ms |
run COCO val2017 vehicle subset (250 images, 971 boxes), 800 px, 1-core x86, run 2026-09-17. the winning config's calibrated thresholds: car , motorcycle , bus , truck ; NMS IoU 0.50. 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
- Apache-2.0 attribution noticesno AGPL anywhere in the tree
- 10,000 hosted API callstop-ups never expire