v3.29: person box shrink 15% + NMS 0.35 to reduce FP

sv_gpu.py

@@ -1,5 +1,5 @@
 """
-Score Vision SN44 — Unified miner v3.28 (2026-04-08). R9c vehicle FP16 (mAP50=0.929). Person: TTA consensus.
+Score Vision SN44 — Unified miner v3.29 (2026-04-08). R9c vehicle FP16 (mAP50=0.929). Person: TTA consensus + 15% box shrink + NMS 0.35.
 Dual-model: vehicle (YOLO11m INT8 1280) + person (YOLO12s FP16 960 TRT).
 Pose model: YOLOv8n-pose FP16 640 for false-positive filtering + keypoint box refinement.
 Vehicle weights loaded from secondary HF repo (meaculpitt/ScoreVision-Vehicle).
@@ -286,7 +286,7 @@ PER_MAX_AREA_RATIO = 0.80
 PER_TILE_OVERLAP = 0.20          # 20% overlap between tiles
 PER_TILE_MIN_DIM_RATIO = 1.15   # tile when image dim > model_dim * this (~1104px for 960 model)
 PER_TILE_CONF = 0.55            # raised from 0.40 to match PER_CONF_LOW
-PER_NMS_IOU = 0.50              # NMS IoU for merging across passes (max-conf wins)
+PER_NMS_IOU = 0.35              # NMS IoU for merging across passes — tightened to reduce FP duplicates
 PER_MAX_DET = 100               # Loose safety ceiling ONLY — not a count cap. Strategy is confidence-floor:
                                 # PER_CONF_LOW=0.60 is the real filter; any box above threshold passes.
                                 # Raised from 50 after 2026-04-05 investigation: top peers emit 77+ boxes on
@@ -2024,11 +2024,16 @@ class Miner:
             if area / img_area > PER_MAX_AREA_RATIO:
                 continue
             b = merged_b[i]
+            # Shrink box 15% toward center to tighten fit (our boxes avg 57% larger than top miners')
+            cx = (b[0] + b[2]) / 2.0
+            cy = (b[1] + b[3]) / 2.0
+            bw2 = (b[2] - b[0]) * 0.85 / 2.0
+            bh2 = (b[3] - b[1]) * 0.85 / 2.0
             out.append(BoundingBox(
-                x1=max(0, min(ow, int(b[0]))),
-                y1=max(0, min(oh, int(b[1]))),
-                x2=max(0, min(ow, int(b[2]))),
-                y2=max(0, min(oh, int(b[3]))),
+                x1=max(0, min(ow, int(cx - bw2))),
+                y1=max(0, min(oh, int(cy - bh2))),
+                x2=max(0, min(ow, int(cx + bw2))),
+                y2=max(0, min(oh, int(cy + bh2))),
                 cls_id=0,
                 conf=max(0.0, min(1.0, float(merged_s[i]))),
             ))