petrol-v8 β the absolute solution
Single 22.7 MB ONNX (havengo's proven-best weights), but layered with the maximum-information preprocessing + TTA pipeline:
- CLAHE on the L-channel (rescues low-contrast / dusk frames)
- Mean-color letterbox padding (vs grey 114) β fewer edge false positives
- 4-pass TTA: 2 scales (1280, 1408) Γ {original, h-flip}
- WBF fusion across all 4 outputs (no agreement gate β keeps singletons)
- Class-aware confidence thresholds (pumps lower, canopy higher)
- Geometric sanity filter (drops 0px or full-image junk)
Designed under the constraint that the honest local benchmark proved filtering moves (agreement gates, conf thresholds post-WBF) HURT scores. v8 only adds techniques that contribute information.
Repo size: 22.7 MB. Expected p95 latency: ~6 s (under 10 s gate).
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support