birds.

detr-r50-fp32 — calibrated

F1 0.604

same checkpoint — stock

F1 0.479

delta +0.124 F1 — COCO val2017 birds subset (125 images, 427 boxes), 800 px letterbox, 1-core x86, run 2026-10-01.

Bird detection — wildlife survey, birdstrike logging, counting flocks on aerial imagery. This is the hardest domain we have measured yet: small, fast objects against open sky. We raced every Apache-2.0 architecture on a held-out bird set, calibrated the winner, and ship it as a pack with the numbers attached — including the honest ceiling of what stock models can do here.

The run, on a real frame.

This is the winning detector, unedited output: one bird — a red kite filling the frame — marked at the shipped operating point. The image was not uploaded anywhere — the box below was 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.

run COCO val2017 birds subset (125 images, 427 boxes), 800 px letterbox, 1-core x86, run 2026-10-01. the winning config's calibrated thresholds: bird 0.90 (COCO class id 16); 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 birds subset (125 images, 427 boxes)
• Apache-2.0 attribution noticesno AGPL anywhere in the tree
• 10,000 hosted API callstop-ups never expire