scorevision: push artifact

README.md

@@ -1,11 +1,16 @@
----
-tags:
-- element_type:detect
-- object:person
-manako:
-  description: Person detection — trained for TurboVision manak0/Detect-Person.
-  input_payload:
-  - {name: frame, type: image, description: RGB frame}
-  output_payload:
-  - {name: detections, type: detections, description: List of detections}
----
+# Petrol-Station Detection Miner — SN44 (TurboVision)
+
+Element: `manak0_Detect-petrol-station-1-0`
+Backbone: YOLO11s, NMS-baked ONNX, FP16 weights (~19.3 MB).
+Classes: `petrol hose`, `petrol pump`, `price board`, `roof canopy`.
+
+## Files
+- `miner.py` — chute entrypoint (`Miner.predict_batch`).
+- `best_fp16.onnx` — model weights, NMS embedded.
+- `class_names.txt` — class id → name mapping.
+- `chute_config.yml` — Chutes runtime spec (Pro 6000, ORT-CUDA).
+
+## Inference
+- Letterbox 1280×1280, BGR→RGB, /255, NCHW float16.
+- Single ORT CUDA pass + horizontal-flip TTA, merged via per-class hard NMS.
+- Per-class confidence thresholds tuned on val.