petrol-fastwin-v1
Fast-win post-processing tweaks layered on top of the havengo/petrol_go1
ONNX baseline:
- conf threshold = 0.50 (matches parent)
- Gaussian Soft-NMS (σ=0.5) replacing hard per-class NMS
- soft-NMS score floor = conf_thres (0.50) — preserves precision
- per-image mean-color letterbox padding
Same weights, smarter inference. No retraining required.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support