s3u1d4o0/vision1

car_wash_miner

Your trained car-wash detection miner (element manak0/Detect-car-wash).

• weights.onnx — YOLO11s, 1280×1280, raw output (NMS done in miner.py). ~19 MB (≤ 30 MB cap). Trained val: mAP50 ≈ 0.776 (nozzle is the weak class — recall ≈ 0.55).
• miner.py — inference entrypoint (copied from cw3): dtype-adaptive input, reads ONNX class metadata, per-class confidence thresholds + per-class NMS. class_names are hardcoded to [broom, drainage gate, nozzle, track].
• class_names.txt, chute_config.yml — deploy metadata.
• 1_infer.py, 2_to_labelstudio.py, 3_extract_labels.py — the inference→Label Studio helpers (LS model_version is tagged car_wash_miner).

Compare against cw3 in Label Studio

Run THIS miner over the same less frames cw3 was run on, then diff the two LS sets.

cd data_top_miner_model/car_wash_miner

# 1) infer with this model (--repo defaults to this dir → loads ./weights.onnx)
python 1_infer.py ../../dataset/car-wash_top_2/less \
    --output-dir ../../dataset/car-wash_top_2/1_less_infer_carwash

# 2) build Label Studio tasks (model_version = car_wash_miner)
python 2_to_labelstudio.py ../../dataset/car-wash_top_2/1_less_infer_carwash \
    --embed-image -o ../../dataset/car-wash_top_2/2_less_label_studio_carwash

Import 2_less_label_studio_carwash/ next to cw3's 2_less_label_studio/ in Label Studio and compare predictions per frame. Note: miner.py's per-class confidence thresholds (_conf_thres_array = [0.37, 0.30, 0.37, 0.45]) are cw3-tuned — adjust them for this model if needed.

Deploy

sv -v deploy-os-miner --model-path data_top_miner_model/car_wash_miner \
    --element-id manak0/Detect-car-wash