Merge pull request #1192 from Justin900429/master

Improvement for BoT-SORT

boxmot/motion/cmc/sof.py

@@ -118,20 +118,16 @@ def apply(self, img, dets):
                 return H
 
         # sparse otical flow for sparse features using Lucas-Kanade with pyramids
+        # calculate new positions of the keypoints between the previous frame (self.prev_img)
+        # and the current frame (img) using sparse optical flow (Lucas-Kanade with pyramids)
         try:
-            matchedKeypoints, status, err = cv2.calcOpticalFlowPyrLK(
+            next_keypoints, status, err = cv2.calcOpticalFlowPyrLK(
                 self.prev_img, img, self.prev_keypoints, None
             )
         except Exception as e:
             LOGGER.warning(f'calcOpticalFlowPyrLK failed: {e}')
             return H
 
-        # calculate new positions of the keypoints between the previous frame (self.prev_img)
-        # and the current frame (img) using sparse optical flow (Lucas-Kanade with pyramids)
-        next_keypoints, status, err = cv2.calcOpticalFlowPyrLK(
-            self.prev_img, img, self.prev_keypoints, None
-        )
-
         # for simplicity, if no keypoints are found, we discard the frame
         if next_keypoints is None:
             return H

boxmot/trackers/botsort/bot_sort.py

@@ -196,6 +196,7 @@ def __init__(
         appearance_thresh: float = 0.25,
         cmc_method: str = "sparseOptFlow",
         frame_rate=30,
+        fuse_first_associate: bool = False,
     ):
         self.tracked_stracks = []  # type: list[STrack]
         self.lost_stracks = []  # type: list[STrack]
@@ -222,6 +223,7 @@ def __init__(
         )
 
         self.cmc = SparseOptFlow()
+        self.fuse_first_associate = fuse_first_associate
 
     def update(self, dets, img):
         assert isinstance(
@@ -288,6 +290,8 @@ def update(self, dets, img):
         # Associate with high score detection boxes
         ious_dists = iou_distance(strack_pool, detections)
         ious_dists_mask = ious_dists > self.proximity_thresh
+        if self.fuse_first_associate:
+          ious_dists = fuse_score(ious_dists, detections)
 
         emb_dists = embedding_distance(strack_pool, detections) / 2.0
         emb_dists[emb_dists > self.appearance_thresh] = 1.0