add conf_cost_weight to stage-1 assignment (default=0.0 disabled)

- Confidence boost in Hungarian cost: solver_iou *= (1 + w * conf[det])
- Neutral at all tested defaults (0.0–0.5); added to Optuna search space [0.0, 1.0]
- IDSW improved 297→293 at w=0.3 but HOTA regressed; w=0.1 exactly neutral

---
Co-authored-by: Claude Code <noreply@anthropic.com>

Author: Borda

autotrack/default_config.json

@@ -16,7 +16,8 @@
         "q_miss_alpha": 0.282,
         "max_interpolation_gap": 32,
         "p_reset_threshold": 26,
-        "oru_threshold": 14
+        "oru_threshold": 14,
+        "conf_cost_weight": 0.0
     },
     "sort": {
         "lost_track_buffer": 30,
@@ -34,4 +35,4 @@
         "delta_t": 3,
         "max_interpolation_gap": 0
     }
-}
+}

autotrack/optimize_tracking.py

@@ -185,6 +185,7 @@ def _build_tracker(params: dict, tracker_name: str):
             stage2_iou_threshold=params["stage2_iou_threshold"],
             iou_age_weight=params["iou_age_weight"],
             high_conf_det_threshold=params["high_conf_det_threshold"],
+            conf_cost_weight=params.get("conf_cost_weight", 0.0),
         )
     if tracker_name == "sort":
         from trackers import SORTTracker

autotrack/search_space.json

@@ -87,6 +87,11 @@
             "type": "int",
             "low": 0,
             "high": 15
+        },
+        "conf_cost_weight": {
+            "type": "float",
+            "low": 0.0,
+            "high": 1.0
         }
     },
     "sort": {

trackers/core/bytetrack/tracker.py

@@ -82,6 +82,7 @@ def __init__(
         stage2_iou_threshold: float = 0.05,
         iou_age_weight: float = 0.03,
         high_conf_det_threshold: float = 0.6,
+        conf_cost_weight: float = 0.0,
     ) -> None:
         # Calculate maximum frames without update based on lost_track_buffer and
         # frame_rate. This scales the buffer based on the frame rate to ensure
@@ -93,6 +94,7 @@ def __init__(
         self.iou_age_weight = iou_age_weight
         self.track_activation_threshold = track_activation_threshold
         self.high_conf_det_threshold = high_conf_det_threshold
+        self.conf_cost_weight = conf_cost_weight
         self.tracks: list[ByteTrackKalmanBoxTracker] = []
 
     def update(
@@ -162,6 +164,14 @@ def update(
         else:
             solver_iou = iou_matrix
 
+        # Confidence boost: scale up solver IoU for higher-confidence detections
+        # so the Hungarian assignment prefers confident detections over uncertain
+        # ones when IoU values are close.  The boost only affects ranking; the
+        # threshold gate still uses raw IoU so valid matches are never blocked.
+        if self.conf_cost_weight > 0 and solver_iou.size > 0 and len(high_indices) > 0:
+            conf_boost = 1.0 + self.conf_cost_weight * confidences[high_indices]
+            solver_iou = solver_iou * conf_boost[np.newaxis, :]
+
         matched, unmatched_tracks, unmatched_high = self._get_associated_indices(
             solver_iou, self.minimum_iou_threshold, raw_similarity=iou_matrix
         )
@@ -259,7 +269,9 @@ def _get_associated_indices(
         unmatched_detections = set(range(n_detections))
 
         # Use raw similarity for threshold gating when available
-        thresh_matrix = raw_similarity if raw_similarity is not None else similarity_matrix
+        thresh_matrix = (
+            raw_similarity if raw_similarity is not None else similarity_matrix
+        )
 
         if n_tracks > 0 and n_detections > 0:
             row_indices, col_indices = linear_sum_assignment(