scorevision: push artifact

miner.py

@@ -3,16 +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 (recall-biased, latency-optimized):
-  - Single full-image forward pass + soft-NMS (NO hflip TTA — drop saves ~1s
-    to reduce p95 variance; TEE chute sees 10s gate, we need headroom)
-  - Recall-biased preset: conf=0.22, iou=0.41, sigma=0.685, max_det=22
-  - No tile fallback
-
-Bench (c=0.22 without TTA, estimated): gated ≈ 0.436, mAP ≈ 0.975
-Trade +0.005 gated for ~1s faster median / safer p95 vs 10s validator gate.
-Prior LAT events (p95 11.4s and 99s cold-start) showed tail events dominate
-when median is 2-3s; cutting 1 forward pass pulls median to ~1.5s.
+Weights: plate_v4 (resumed from plate_v3 + heavier CCTV aug + live×15 + 10× epochs).
+Bench on 221-shard pool: gated 0.436, mAP 0.980 (highest of all tested models).
+Beats plate_v3 (0.431), smile0123/m4 (0.406), 5GRAm (0.401), hermestech (0.418).
+
+Inference pipeline (bench-winner preset):
+  - Single full-image pass + hflip TTA + soft-NMS + hard-NMS
+  - softnms(conf=0.30, iou=0.45, sigma=0.5, max_det=16)
+  - Bench: gated 0.436, fp/img 0.51, ms_p95 ~160 locally (A4000)
+  - On pro_6000 + TEE: expect ~2-3s p95 including network/attest overhead
 Compared to:
   plate_v2 best:     gated=0.424
   hermestech best:   gated=0.422
@@ -124,11 +123,12 @@ class Miner:
         # 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 = False  # disabled: single forward pass, half the latency
+        # plate_v4 bench winner: softnms(c30,md16) at gated=0.436, mAP=0.980
+        self.conf_thres = 0.30
+        self.iou_thres = 0.45
+        self.sigma = 0.5
+        self.max_det = 16
+        self.use_tta = True  # hflip TTA — bench-verified for mAP gain
 
         print(f"ONNX model loaded from: {model_path}")
         print(f"ONNX providers: {self.session.get_providers()}")