fix bugs in c_biou track

Author: JackWoo0831

README.md

@@ -21,6 +21,7 @@ git checkout v2  # change to v2 branch !!
 
 ## 🗺️ Latest News
 
+- ***2024.11.29*** Fix bugs of C-BIoU Track (the state prediction and updating bugs)
 - ***2024.10.24*** Add Hybrid SORT and fix some errors and bugs of OC-SORT.
 
 ## ❤️ Introduction

README_CN.md

@@ -15,6 +15,7 @@ git checkout v2  # change to v2 branch !!
 
 ## 🗺️ 最近更新
 
+- ***2024.11.29*** 修复了C-BIoU Tracker中轨迹状态的更新和预测的错误
 - ***2024.10.24*** 增加了 Hybrid SORT 并且修复了OC-SORT的一些bug和错误。
 
 

tracker/trackers/c_biou_tracker.py

@@ -39,18 +39,11 @@ def update(self, output_results, img, ori_img):
         categories = output_results[:, -1]
 
         remain_inds = scores > self.args.conf_thresh
-        inds_low = scores > 0.1
-        inds_high = scores < self.args.conf_thresh
-
-        inds_second = np.logical_and(inds_low, inds_high)
-        dets_second = bboxes[inds_second]
         dets = bboxes[remain_inds]
 
         cates = categories[remain_inds]
-        cates_second = categories[inds_second]
 
         scores_keep = scores[remain_inds]
-        scores_second = scores[inds_second]
 
         if len(dets) > 0:
             '''Detections'''
@@ -68,7 +61,7 @@ def update(self, output_results, img, ori_img):
             else:
                 tracked_tracklets.append(track)
 
-        ''' Step 2: First association, with high score detection boxes'''
+        ''' Step 2: First association, with small buffer IoU'''
         tracklet_pool = joint_tracklets(tracked_tracklets, self.lost_tracklets)
 
         # Predict the current location with Kalman
@@ -89,40 +82,29 @@ def update(self, output_results, img, ori_img):
                 track.re_activate(det, self.frame_id, new_id=False)
                 refind_tracklets.append(track)
 
-        ''' Step 3: Second association, with low score detection boxes'''
-        # association the untrack to the low score detections
-        if len(dets_second) > 0:
-            '''Detections'''
-            detections_second = [Tracklet_w_bbox_buffer(tlwh, s, cate, motion=self.motion) for
-                          (tlwh, s, cate) in zip(dets_second, scores_second, cates_second)]
-        else:
-            detections_second = []
-        r_tracked_tracklets = [tracklet_pool[i] for i in u_track if tracklet_pool[i].state == TrackState.Tracked]
+        unmatched_tracklets = [tracklet_pool[i] for i in u_track if tracklet_pool[i].state == TrackState.Tracked]
+        unmatched_detections = [detections[i] for i in u_detection]
 
+        '''Step 3: Second association, with large buffer IoU'''
 
-        dists = buffered_iou_distance(r_tracked_tracklets, detections_second, level=2)
+        dists = buffered_iou_distance(unmatched_tracklets, unmatched_detections, level=2)
+
+        matches, u_track, u_detection = linear_assignment(dists, thresh=0.5)
 
-        matches, u_track, u_detection_second = linear_assignment(dists, thresh=0.5)
         for itracked, idet in matches:
-            track = r_tracked_tracklets[itracked]
-            det = detections_second[idet]
+            track = unmatched_tracklets[itracked]
+            det = unmatched_detections[idet]
             if track.state == TrackState.Tracked:
                 track.update(det, self.frame_id)
                 activated_tracklets.append(track)
             else:
                 track.re_activate(det, self.frame_id, new_id=False)
                 refind_tracklets.append(track)
 
-        for it in u_track:
-            track = r_tracked_tracklets[it]
-            if not track.state == TrackState.Lost:
-                track.mark_lost()
-                lost_tracklets.append(track)
-
         '''Deal with unconfirmed tracks, usually tracks with only one beginning frame'''
-        detections = [detections[i] for i in u_detection]
+        detections = [unmatched_detections[i] for i in u_detection]
         dists = buffered_iou_distance(unconfirmed, detections, level=1)
-       
+
         matches, u_unconfirmed, u_detection = linear_assignment(dists, thresh=0.7)
 
         for itracked, idet in matches:
@@ -133,15 +115,15 @@ def update(self, output_results, img, ori_img):
             track.mark_removed()
             removed_tracklets.append(track)
 
-        """ Step 4: Init new tracklets"""
+        '''Step 4. Inital new tracks'''
         for inew in u_detection:
             track = detections[inew]
             if track.score < self.det_thresh:
                 continue
             track.activate(self.frame_id)
             activated_tracklets.append(track)
 
-        """ Step 5: Update state"""
+        ''' Step 5: Update state'''
         for track in self.lost_tracklets:
             if self.frame_id - track.end_frame > self.max_time_lost:
                 track.mark_removed()

tracker/trackers/tracklet.py

@@ -438,7 +438,14 @@ class Tracklet_w_bbox_buffer(Tracklet):
     Tracklet class with buffer of bbox, for C_BIoU track.
     """
     def __init__(self, tlwh, score, category, motion='byte'):
-        super().__init__(tlwh, score, category, motion)
+        # initial position
+        self._tlwh = np.asarray(tlwh, dtype=np.float)
+        self.is_activated = False
+
+        self.score = score
+        self.category = category
+
+        # Note in C-BIoU tracker the kalman filter is abandoned
 
         # params in motion state
         self.b1, self.b2, self.n = 0.3, 0.5, 5
@@ -448,6 +455,7 @@ def __init__(self, tlwh, score, category, motion='byte'):
         self.buffer_bbox1 = self.get_buffer_bbox(level=1)
         self.buffer_bbox2 = self.get_buffer_bbox(level=2)
         # motion state, s^{t + \delta} = o^t + (\delta / n) * \sum_{i=t-n+1}^t(o^i - o^{i-1}) = o^t + (\delta / n) * (o^t - o^{t - n})
+        self.motion_state0 = self._tlwh  # original tlwh
         self.motion_state1 = self.buffer_bbox1.copy()
         self.motion_state2 = self.buffer_bbox2.copy()
 
@@ -483,22 +491,38 @@ def re_activate(self, new_track, frame_id, new_id=False):
         # update stored bbox
         if (len(self.origin_bbox_buffer) > self.n):
             self.origin_bbox_buffer.popleft()
-            self.origin_bbox_buffer.append(self._tlwh)
-        else:
-            self.origin_bbox_buffer.append(self._tlwh)
+
+        self.origin_bbox_buffer.append(self._tlwh)
 
         self.buffer_bbox1 = self.get_buffer_bbox(level=1)
         self.buffer_bbox2 = self.get_buffer_bbox(level=2)
+
+        self.motion_state0 = self._tlwh
         self.motion_state1 = self.buffer_bbox1.copy()
         self.motion_state2 = self.buffer_bbox2.copy()
 
+    def predict(self):
+        # Note that in C-BIoU Tracker, no need to use Kalman Filter
+        self.time_since_update += 1
+
+        # Average motion model: s^{t + \delta} = o^t + (\delta / n) * (o^t - o^{t - n})
+        assert len(self.origin_bbox_buffer), 'The bbox buffer is empty'
+
+        motion_state = self.origin_bbox_buffer[-1] + \
+                    (self.time_since_update / len(self.origin_bbox_buffer)) * (self.origin_bbox_buffer[-1] - self.origin_bbox_buffer[0])
+        
+        self.motion_state0 = motion_state
+        self.motion_state1 = self.get_buffer_bbox(level=1, bbox=motion_state)
+        self.motion_state2 = self.get_buffer_bbox(level=2, bbox=motion_state)
+
+
     def update(self, new_track, frame_id):
         self.frame_id = frame_id
 
         new_tlwh = new_track.tlwh
         self.score = new_track.score
 
-        self.kalman_filter.update(self.convert_func(new_tlwh))
+        # self.kalman_filter.update(self.convert_func(new_tlwh))  # no need to use Kalman Filter
 
         self.state = TrackState.Tracked
         self.is_activated = True
@@ -508,25 +532,15 @@ def update(self, new_track, frame_id):
         # update stored bbox
         if (len(self.origin_bbox_buffer) > self.n):
             self.origin_bbox_buffer.popleft()
-            self.origin_bbox_buffer.append(new_tlwh)
-        else:
-            self.origin_bbox_buffer.append(new_tlwh)
-
-        # update motion state
-        if self.time_since_update:  # have some unmatched frames
-            if len(self.origin_bbox_buffer) < self.n:
-                self.motion_state1 = self.get_buffer_bbox(level=1, bbox=new_tlwh)
-                self.motion_state2 = self.get_buffer_bbox(level=2, bbox=new_tlwh)
-            else:  # s^{t + \delta} = o^t + (\delta / n) * (o^t - o^{t - n})
-                motion_state = self.origin_bbox_buffer[-1] + \
-                    (self.time_since_update / self.n) * (self.origin_bbox_buffer[-1] - self.origin_bbox_buffer[0])
-                self.motion_state1 = self.get_buffer_bbox(level=1, bbox=motion_state)
-                self.motion_state2 = self.get_buffer_bbox(level=2, bbox=motion_state)
-
-        else:  # no unmatched frames, use current detection as motion state
-            self.motion_state1 = self.get_buffer_bbox(level=1, bbox=new_tlwh)
-            self.motion_state2 = self.get_buffer_bbox(level=2, bbox=new_tlwh)
+        self.origin_bbox_buffer.append(new_tlwh)
 
+    # Drop kalman filter, rewrite the tlwh function
+    @property
+    def tlwh(self):
+        """Get current position in bounding box format `(top left x, top left y,
+                width, height)`.
+        """
+        return self.motion_state0
 
 class Tracklet_w_depth(Tracklet):
     """