scorevision: push artifact

miner.py

@@ -189,7 +189,7 @@ VEH_NMS_IOU = 0.50
 VEH_CLASS_CONF: dict[int, float] = {
     1: 0.45,    # car — most FP-prone class (75% of training data, overconfident)
     2: 0.45,    # truck — moderate raise
-    3: 0.35,    # motorcycle — keep lower (small targets, easy to miss)
+    3: 0.45,    # motorcycle — raised from 0.35 to reduce FP (small targets, easy to miss)
     4: 1.0,     # bus — effectively suppressed (not scored anyway)
 }
 
@@ -215,7 +215,7 @@ VEH_MIN_WH = 8
 VEH_MIN_AREA = 100
 VEH_MAX_ASPECT = 8.0
 VEH_MAX_AREA_RATIO = 0.95
-VEH_MAX_DET = 150
+VEH_MAX_DET = 40
 
 # ── Vehicle parts confirmation config ────────────────────────────────────
 # Cross-validates vehicle detections using person detections, OpenCV analysis,
@@ -245,7 +245,7 @@ VEH_PARTS_PLATE_MIN_PX = 120  # only check plates on medium+ vehicles      # Min
 VEH_PARTS_PLATE_CONF = 0.35      # Min plate detection confidence
 
 # ── Person config (TTA consensus) ───────────────────────────────────────────
-PER_CONF_LOW = 0.45
+PER_CONF_LOW = 0.55
 PER_CONF_HIGH = 0.58
 PER_CONSENSUS_IOU = 0.50
 PER_RTF_BUDGET = 8.0
@@ -259,8 +259,9 @@ PER_MAX_AREA_RATIO = 0.80
 # ── Person tiling config (SAHI-inspired) ────────────────────────────────────
 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 = 0.40            # lower threshold for tile passes (NMS handles FP)
+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_MAX_DET = 15                # hard cap on person detections per image
 
 # ── Pose FP filter + box refinement config ──────────────────────────────────
 POSE_CONF_THRESH = 0.25         # Minimum confidence for pose detection
@@ -1703,13 +1704,12 @@ class Miner:
     def _infer_person(self, image_bgr):
         """Person detection with SAHI-inspired tiled inference + dynamic NMS.
 
-        Pipeline:
+        Pipeline (2-pass, optimized for RTF):
         1. Full-image pass at native 960px
-        2. 2 tiled passes (higher effective resolution for small/distant people)
-        3. Flip TTA pass if time budget allows
-        4. Dynamic NMS merge (adapts IoU threshold to scene density)
-        5. Sanity filters
-        6. Pose FP filter + box refinement
+        2. Flip TTA pass
+        3. Dynamic NMS merge (adapts IoU threshold to scene density)
+        4. Sanity filters + PER_MAX_DET cap
+        5. Pose FP filter + box refinement (if time allows)
         """
         oh, ow = image_bgr.shape[:2]
         t_start = time.monotonic()
@@ -1726,35 +1726,14 @@ class Miner:
 
         elapsed_pass1 = time.monotonic() - t_start
 
-        # Pass 2-3: tiled passes
-        tiles = self._generate_tiles(oh, ow)
-        if len(tiles) > 1 and elapsed_pass1 < PER_RTF_BUDGET / 4:
-            for tile_region in tiles[1:]:
-                if time.monotonic() - t_start > PER_RTF_BUDGET * 0.6:
-                    break
-                boxes_t, confs_t = self._per_run_tile(
-                    image_bgr, tile_region, PER_TILE_CONF)
-                if len(boxes_t) > 0:
-                    all_boxes.append(boxes_t)
-                    all_confs.append(confs_t)
-
-        # Pass 4: flip TTA if time allows
-        if time.monotonic() - t_start < PER_RTF_BUDGET / 4:
-            flipped = cv2.flip(image_bgr, 1)
-            boxes_flip, confs_flip = self._per_run_pass(flipped, PER_CONF_LOW)
-            if len(boxes_flip) > 0:
-                boxes_flip[:, 0], boxes_flip[:, 2] = (
-                    ow - boxes_flip[:, 2], ow - boxes_flip[:, 0])
-                all_boxes.append(boxes_flip)
-                all_confs.append(confs_flip)
-
-        # Pass 5: CLAHE enhanced preprocessing pass
-        if time.monotonic() - t_start < PER_RTF_BUDGET / 4:
-            enhanced = self._per_enhance(image_bgr)
-            boxes_enh, confs_enh = self._per_run_pass(enhanced, PER_CONF_LOW)
-            if len(boxes_enh) > 0:
-                all_boxes.append(boxes_enh)
-                all_confs.append(confs_enh)
+        # Pass 2: flip TTA (always run — only 2 passes total for RTF safety)
+        flipped = cv2.flip(image_bgr, 1)
+        boxes_flip, confs_flip = self._per_run_pass(flipped, PER_CONF_LOW)
+        if len(boxes_flip) > 0:
+            boxes_flip[:, 0], boxes_flip[:, 2] = (
+                ow - boxes_flip[:, 2], ow - boxes_flip[:, 0])
+            all_boxes.append(boxes_flip)
+            all_confs.append(confs_flip)
 
         if not all_boxes:
             return []
@@ -1766,6 +1745,12 @@ class Miner:
         nms_iou = 0.60 if n_raw > 30 else (0.40 if n_raw < 10 else PER_NMS_IOU)
         merged_b, merged_s = self._nms_max_conf(merged_b, merged_s, nms_iou)
 
+        # Hard cap on max detections (FP protection)
+        if len(merged_s) > PER_MAX_DET:
+            top_idx = np.argsort(merged_s)[-PER_MAX_DET:]
+            merged_b = merged_b[top_idx]
+            merged_s = merged_s[top_idx]
+
         if len(merged_b) == 0:
             return []