SN44 Detect-car-wash objects miner — armN winner-supervised (ONNX, 704px)
Model: YOLO11n detector (Ultralytics), trained on official-domain challenge frames
labeled by winner-consensus (map50>=0.9, fp>=0.9 record boxes) — frozen-GT-proxy
supervision matching the official sam3_json_v1 scoring target.
Winner-proxy holdout (120 frames, disjoint): map50=0.8364, fp=0.951, gated=0.4889
(internal bar 0.45; leader rolling 0.497). 1024-parent gate = 0.4987.
Latency: fp32 ONNX, 85ms p95 single-frame on 2 vCPU-class cores (measured 2026-08-22)
— within the 100ms+10ms compliance budget at 704px inference.
Packaging: CPU-onnxruntime, no network access, no env reads, no symlinks.
Channel-order contract (measured)
Input arrives as cv2.imdecode(IMREAD_COLOR) => BGR. _predict_one converts BGR->RGB before
preprocess (measured cost of NOT converting: map50 0.8364 -> 0.6307 on the 120-frame winner-proxy holdout).
