scorevision: push artifact

chute_config.yml

@@ -13,3 +13,4 @@ Chute:
   concurrency: 4
   max_instances: 5
   scaling_threshold: 0.5
+  shutdown_after_seconds: 288000

miner.py

@@ -178,10 +178,12 @@ VEH_CONF_DEFAULT = 0.35
 VEH_TTA_CONF = 0.10
 VEH_WBF_IOU = 0.40
 
-# ── Person config ───────────────────────────────────────────────────────────
-PER_CONF = 0.40
-PER_TTA_CONF = 0.25
-PER_WBF_IOU = 0.40
+# ── Person config (TTA consensus) ───────────────────────────────────────────
+PER_CONF_LOW = 0.32          # Detection threshold (both passes)
+PER_CONF_HIGH = 0.55         # Above this = unconditionally kept
+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)
 
 # ── Shared ──────────────────────────────────────────────────────────────────
 WBF_SKIP_THR = 0.0001
@@ -508,36 +510,71 @@ class Miner:
 
     def _infer_person(self, image_bgr):
         oh, ow = image_bgr.shape[:2]
-        all_b, all_s = [], []
 
-        def _collect(boxes, confs):
-            if len(boxes) == 0:
-                return
-            norm = boxes.copy()
-            norm[:, [0, 2]] /= ow
-            norm[:, [1, 3]] /= oh
-            norm = np.clip(norm, 0, 1)
-            all_b.append(norm)
-            all_s.append(confs)
+        # Pass 1: original
+        t_start = time.monotonic()
+        boxes_orig, confs_orig = self._per_run_pass(image_bgr, PER_CONF_LOW)
+        elapsed = time.monotonic() - t_start
 
-        if ENABLE_TTA:
-            # Pass 1: original
-            _collect(*self._per_run_pass(image_bgr, PER_TTA_CONF))
-            # Pass 2: hflip
+        # Pass 2: hflip — only if RTF budget allows
+        has_flip = False
+        if elapsed < PER_RTF_BUDGET / 2:
             flipped = cv2.flip(image_bgr, 1)
-            bx, sc = self._per_run_pass(flipped, PER_TTA_CONF)
-            if len(bx):
-                bx[:, 0], bx[:, 2] = ow - bx[:, 2], ow - bx[:, 0]
-            _collect(bx, sc)
+            boxes_flip, confs_flip = self._per_run_pass(flipped, PER_CONF_LOW)
+            if len(boxes_flip):
+                boxes_flip[:, 0], boxes_flip[:, 2] = ow - boxes_flip[:, 2], ow - boxes_flip[:, 0]
+            has_flip = True
         else:
-            # Single pass
-            _collect(*self._per_run_pass(image_bgr, PER_CONF))
+            logger.warning(f"[person TTA] skipping flip — pass1 took {elapsed*1000:.0f}ms")
+
+        # ── Consensus filtering ─────────────────────────────────────────
+        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))
 
-        if not all_b:
+            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])
+
+            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 not kept_b:
             return []
 
-        # Always apply WBF — deduplicates overlapping anchors from raw ONNX output
-        fb, fs = _wbf_single(all_b, all_s, iou_thr=PER_WBF_IOU, skip_thr=WBF_SKIP_THR)
+        # ── WBF dedup (merges near-duplicate detections) ────────────────
+        kept_b = np.array(kept_b)
+        kept_s = np.array(kept_s)
+        norm = kept_b.copy()
+        norm[:, [0, 2]] /= ow
+        norm[:, [1, 3]] /= oh
+        norm = np.clip(norm, 0, 1)
+
+        fb, fs = _wbf_single([norm], [kept_s], iou_thr=PER_WBF_IOU, skip_thr=WBF_SKIP_THR)
 
         if len(fb) == 0:
             return []
@@ -545,9 +582,6 @@ class Miner:
         fb[:, [0, 2]] *= ow
         fb[:, [1, 3]] *= oh
 
-        keep = fs >= PER_CONF
-        fb, fs = fb[keep], fs[keep]
-
         out = []
         for i in range(len(fb)):
             b = fb[i]
@@ -561,6 +595,7 @@ class Miner:
             ))
         return out
 
+    
     # ── Unified inference ───────────────────────────────────────────────────
 
     def _infer_single(self, image_bgr: ndarray) -> list[BoundingBox]:

vehicle_weights.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:14045470e36fe94179154a9127b4d8e21228e87a801ac078a2fe77477d5f3c8b
-size 19272252
+oid sha256:00bae032cee689d04b3b9131cd80134d03c17972490190f45a5f2aa96f9b703a
+size 21244589