ep29 weights + miner.py pipeline tuning (NMS sigma 0.3, SAHI overlap 5%, no adaptive fallback)

miner.py

@@ -163,7 +163,9 @@ class Miner:
         # given 0.6 weight on map50 vs 0.4 on false_positive in composite.
         self.conf_threshold = 0.16
         # Soft-NMS hyperparameters (Gaussian variant).
-        self.soft_nms_sigma = 0.5
+        # Tightened 0.5 → 0.3: sharper decay collapses near-duplicates from SAHI
+        # tile-seam overlaps faster, dropping more below the 0.01 score floor.
+        self.soft_nms_sigma = 0.3
         # Final score floor after Soft-NMS decay. Was 0.20 — raised threshold
         # killed decayed real plates (e.g. plate adjacent to a higher-conf
         # detection gets decayed below 0.20 and dropped). Matches competitor
@@ -445,8 +447,8 @@ class Miner:
         """Run the quad-4 tile pipeline and return RAW (pre-Soft-NMS)
         detections in original-image coordinates."""
         orig_h, orig_w = image_bgr.shape[:2]
-        OVERLAP_X = 70   # ~10% of 1408/2
-        OVERLAP_Y = 38   # ~10% of 768/2
+        OVERLAP_X = 35   # ~5% of 1408/2 (was 70/10%; halved to cut tile-seam duplicates)
+        OVERLAP_Y = 19   # ~5% of 768/2  (was 38/10%; halved to cut tile-seam duplicates)
         mx = orig_w // 2
         my = orig_h // 2
 
@@ -485,22 +487,10 @@ class Miner:
 
         all_dets = self._quad4_raw_dets(image_bgr)
 
-        # Adaptive conf fallback: if the quad-4 pass produced nothing, retry
-        # once at a lower pre-NMS threshold. Rescues archived floor-drops
-        # where the model had plate signal but nothing crossed 0.18 anywhere
-        # in the 4 tiles (observed on validator tasks 57803/57836/57848).
-        if not all_dets:
-            _orig_conf = self.conf_threshold
-            try:
-                self.conf_threshold = 0.10
-                all_dets = self._quad4_raw_dets(image_bgr)
-            finally:
-                self.conf_threshold = _orig_conf
-            if all_dets:
-                _cuda_log.warning(
-                    "adaptive conf fallback rescued %d raw dets at conf=0.10",
-                    len(all_dets),
-                )
+        # Adaptive conf fallback removed Apr 26: ep 29 has higher recall than
+        # v3 so the empty-first-pass case is rarer, and when it did fire the
+        # conf=0.10 retry generated phantoms (FP drag) AND added a ~3s
+        # inference pass (latency gate trigger). See mining_history.md.
 
         # Keep flipped-view detections SEPARATE from original, so we can use
         # them as a cross-view verifier (hermes-style gate) later — not just

numberplate_weights.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:75e91a8ee137d2f52eef752527a3b3ff1159c5b4417c5b40f205af090fb0ea00
-size 19579923
+oid sha256:f585a11f502df54f9610520b0cdef3b5414732a61543ab63113d0b0eb1239c41
+size 19761271