petrol-v8b β single-pass max-information
Single 22.7 MB ONNX (havengo's proven-best weights). v8b drops the TTA passes that timed out in v8 and keeps only the information-adding components that fit in single-pass latency:
- CLAHE on the L-channel (rescues low-contrast / dusk frames)
- Mean-color letterbox padding (vs grey 114) β fewer edge false positives
- WBF fusion within single pass (per-class)
- Class-aware confidence thresholds (pumps lower, canopy higher)
- Geometric sanity filter (drops 0px or full-image junk)
The honest local benchmark proved filtering moves (agreement gates) hurt scores. v8b only filters obvious geometric junk and uses information-adding preprocessing.
Repo size: 22.7 MB. Expected p95 latency: ~1.4β1.8 s (well under 10 s gate).
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support