scorevision: push artifact

miner.py

@@ -3,13 +3,15 @@
 Base weights: plate_v3 (YOLO26s fine-tuned on Roboflow-filtered + 10x live pseudo-GT,
 resumed from plate_v2). fp16 end2end ONNX, static 1x3x1280x1280, ~19.4 MB.
 
-Inference pipeline (tuned per bench_v2.py on 184-shard pool):
+Inference pipeline (tuned per bench_v2.py + live observation on first 5 shards):
   - Single full-image pass with soft-NMS + hflip TTA
-  - Tight preset: conf=0.30, iou=0.45, sigma=0.5, max_det=16
-  - No tile fallback (v3's mAP=0.973 is already high enough; tiles only add FPs)
+  - Recall-biased preset: conf=0.22, iou=0.41, sigma=0.685, max_det=22
+  - No tile fallback (v3's recall is already high without tiles)
 
 Bench on 184-shard live pseudo-GT pool (/mnt/shadeform-data/plate_research/live_gt/):
-  gated=0.441  mAP=0.973  fp/img=0.29  ms_med=152  ms_p95=161
+  gated=0.440  mAP=0.978 (highest)  fp/img=0.38  ms_p95=157
+Switched from c30 after live shard be77593656fa: we scored 0.168 (mAP 0.500)
+while competitors hit 0.318 (mAP 0.750) — missed borderline-conf plates.
 Compared to:
   plate_v2 best:     gated=0.424
   hermestech best:   gated=0.422
@@ -114,12 +116,17 @@ class Miner:
         self.input_height = self._safe_dim(self.input_shape[2], default=SIZE)
         self.input_width = self._safe_dim(self.input_shape[3], default=SIZE)
 
-        # Tuned preset for plate_v3 (from bench_v2.py, 184-shard live pool).
-        # Best gated=0.441 AND lowest fp/img=0.29 AND tight ms_p95=161.
-        self.conf_thres = 0.30
-        self.iou_thres = 0.45
-        self.sigma = 0.5
-        self.max_det = 16
+        # Tuned preset for plate_v3 — recall-biased variant.
+        # Bench softnms(c22,i.41,s.685) on 184-shard pool:
+        #   gated=0.440  mAP=0.978 (highest)  fp/img=0.38  ms_p95=157
+        # Switched from conf=0.30 after live data showed the tighter threshold
+        # missed borderline plates on shards where competitors scored 0.318.
+        # Trade: slightly higher fp/img on easy shards (capped by max_det), but
+        # recovers recall on hard shards where it matters most.
+        self.conf_thres = 0.22
+        self.iou_thres = 0.41
+        self.sigma = 0.685
+        self.max_det = 22
         self.use_tta = True
 
         print(f"ONNX model loaded from: {model_path}")