scorevision: push artifact

miner.py

@@ -185,6 +185,12 @@ PER_CONSENSUS_IOU = 0.50     # Cross-view IoU to confirm low-conf box
 PER_WBF_IOU = 0.40           # Final WBF dedup (unchanged)
 PER_RTF_BUDGET = 8.0         # Conservative RTF budget (gate is 10s)
 
+# ── Person box sanity filters (matching top scorers) ────────────────────────
+PER_MIN_WH = 8               # Minimum box width or height (pixels)
+PER_MIN_AREA = 14 * 14       # Minimum box area (pixels^2)
+PER_MAX_ASPECT = 8.0         # Maximum aspect ratio
+PER_MAX_AREA_RATIO = 0.80    # Maximum box area / image area
+
 # ── Shared ──────────────────────────────────────────────────────────────────
 WBF_SKIP_THR = 0.0001
 
@@ -527,41 +533,35 @@ class Miner:
         else:
             logger.warning(f"[person TTA] skipping flip — pass1 took {elapsed*1000:.0f}ms")
 
-        # ── Consensus filtering ─────────────────────────────────────────
+        # ── Consensus filtering (v2 — original-primary, flip supplements) ──
         kept_b, kept_s = [], []
 
-        if has_flip:
-            # High-conf (>=0.55): keep unconditionally from both views
-            # Low-conf (0.32-0.55): keep only if confirmed in other view
-
-            def _has_match(box, others):
-                if len(others) == 0:
-                    return False
-                xx1 = np.maximum(box[0], others[:, 0])
-                yy1 = np.maximum(box[1], others[:, 1])
-                xx2 = np.minimum(box[2], others[:, 2])
-                yy2 = np.minimum(box[3], others[:, 3])
-                inter = np.maximum(0, xx2 - xx1) * np.maximum(0, yy2 - yy1)
-                a1 = (box[2] - box[0]) * (box[3] - box[1])
-                a2 = (others[:, 2] - others[:, 0]) * (others[:, 3] - others[:, 1])
-                iou = inter / (a1 + a2 - inter + 1e-9)
-                return bool(np.any(iou > PER_CONSENSUS_IOU))
-
-            for i in range(len(boxes_orig)):
-                if confs_orig[i] >= PER_CONF_HIGH or _has_match(boxes_orig[i], boxes_flip):
-                    kept_b.append(boxes_orig[i])
-                    kept_s.append(confs_orig[i])
+        # Always keep ALL original boxes at conf >= PER_CONF_LOW
+        for i in range(len(boxes_orig)):
+            kept_b.append(boxes_orig[i])
+            kept_s.append(confs_orig[i])
 
+        if has_flip:
+            # From flip view: only add HIGH-conf boxes that original MISSED
+            # (IoU < 0.3 with all original boxes = truly new detection)
             for i in range(len(boxes_flip)):
-                if confs_flip[i] >= PER_CONF_HIGH or _has_match(boxes_flip[i], boxes_orig):
-                    kept_b.append(boxes_flip[i])
-                    kept_s.append(confs_flip[i])
-        else:
-            # No flip — just use original boxes (same as v6 single-pass)
-            for i in range(len(boxes_orig)):
-                if confs_orig[i] >= PER_CONF_LOW:
-                    kept_b.append(boxes_orig[i])
-                    kept_s.append(confs_orig[i])
+                if confs_flip[i] < PER_CONF_HIGH:
+                    continue
+                # Check if this flip box overlaps any original box
+                if len(boxes_orig) > 0:
+                    xx1 = np.maximum(boxes_flip[i, 0], boxes_orig[:, 0])
+                    yy1 = np.maximum(boxes_flip[i, 1], boxes_orig[:, 1])
+                    xx2 = np.minimum(boxes_flip[i, 2], boxes_orig[:, 2])
+                    yy2 = np.minimum(boxes_flip[i, 3], boxes_orig[:, 3])
+                    inter = np.maximum(0, xx2 - xx1) * np.maximum(0, yy2 - yy1)
+                    a1 = (boxes_flip[i, 2] - boxes_flip[i, 0]) * (boxes_flip[i, 3] - boxes_flip[i, 1])
+                    a2 = (boxes_orig[:, 2] - boxes_orig[:, 0]) * (boxes_orig[:, 3] - boxes_orig[:, 1])
+                    iou = inter / (a1 + a2 - inter + 1e-9)
+                    if np.max(iou) >= 0.3:
+                        continue  # Already detected by original — skip
+                # Truly new high-conf detection from flip
+                kept_b.append(boxes_flip[i])
+                kept_s.append(confs_flip[i])
 
         if not kept_b:
             return []
@@ -582,6 +582,28 @@ class Miner:
         fb[:, [0, 2]] *= ow
         fb[:, [1, 3]] *= oh
 
+        # ── Sanity filter (remove degenerate boxes) ─────────────────────
+        img_area = float(oh * ow)
+        sane = []
+        for i in range(len(fb)):
+            bw = fb[i, 2] - fb[i, 0]
+            bh = fb[i, 3] - fb[i, 1]
+            if bw < PER_MIN_WH or bh < PER_MIN_WH:
+                continue
+            area = bw * bh
+            if area < PER_MIN_AREA:
+                continue
+            if max(bw, bh) / max(min(bw, bh), 1e-6) > PER_MAX_ASPECT:
+                continue
+            if area / img_area > PER_MAX_AREA_RATIO:
+                continue
+            sane.append(i)
+
+        if not sane:
+            return []
+        fb = fb[sane]
+        fs = fs[sane]
+
         out = []
         for i in range(len(fb)):
             b = fb[i]