Buy the pack — $149people.
- detr-r50-fp32 — calibrated
- F1 0.768
- same checkpoint — stock
- F1 0.666
delta +0.103 F1 — COCO val2017 person subset (250 images, 926 boxes), 800 px letterbox, 1-core x86, run 2026-09-28.
Person detection — the most requested COCO class there is. We raced every Apache-2.0 architecture on a 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 frame below.
The run, on a real frame.
This is the winning detector, unedited output: six people 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 | 0.822 | 0.722 | 0.666 | 0.768 | 1606 ms |
| rfdetr-n-fp32 | 0.857 | 0.692 | 0.706 | 0.765 | 325 ms |
| rfdetr-n-int8 | 0.864 | 0.678 | 0.699 | 0.759 | 224 ms |
| detr-r50-int8 | 0.780 | 0.668 | 0.616 | 0.719 | 1332 ms |
run COCO val2017 person subset (250 images, 926 boxes), 800 px letterbox, 1-core x86, run 2026-09-28. the winning config's calibrated thresholds: person 0.90 — keyed by COCO class id as 1:0.90; NMS IoU 0.60. 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 person subset (250 images
- Apache-2.0 attribution noticesno AGPL anywhere in the tree
- 10,000 hosted API callstop-ups never expire