[MOT] update vehicle modelzoo (#4360)

* update mot vehicle modelzoo

* update bdd100k visdrone vehicle  modelzoo

configs/mot/README.md

@@ -191,9 +191,9 @@ If you use a stronger detection model, you can get better results. Each txt is t
 ### FairMOT Results on each val-set of Vehicle category
 |    Dataset      |  input shape |  MOTA  |  IDF1  |  FPS   |  download | config |
 | :-------------| :------- | :----: | :----: | :----: | :-----: |:------: |
-|  BDD100K      | 1088x608 |  34.9 |  39.9  |    -    | [model](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_bdd100k_vehicle.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mot/vehicle/fairmot_dla34_30e_1088x608_bdd100k_vehicle.yml) |
-|  KITTI        | 1088x608 |  53.9 |    -   |    -   |[model](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_kitti_vehicle.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mot/vehicle/fairmot_dla34_30e_1088x608_kitti_vehicle.yml) |
-|  VisDrone     | 1088x608 |  29.8 |   51.3 |    -   | [model](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_visdrone_vehicle.pdparams) | [config](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mot/vehicle/fairmot_dla34_30e_1088x608_visdrone_vehicle.yml) |
+|  BDD100K      | 1088x608 |  43.5 |  50.0  |    -    | [model](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_bdd100k_vehicle.pdparams) | [config](./vehicle/fairmot_dla34_30e_1088x608_bdd100k_vehicle.yml) |
+|  KITTI        | 1088x608 |  82.7 |    -   |    -   |[model](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_kitti_vehicle.pdparams) | [config](./vehicle/fairmot_dla34_30e_1088x608_kitti_vehicle.yml) |
+|  VisDrone     | 1088x608 |  52.1 |   63.3 |    -   | [model](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_visdrone_vehicle.pdparams) | [config](./vehicle/fairmot_dla34_30e_1088x608_visdrone_vehicle.yml) |
 
 ## Dataset Preparation
 

configs/mot/README_cn.md

@@ -191,9 +191,10 @@ wget https://dataset.bj.bcebos.com/mot/det_results_dir.zip
 
 |    数据集      |  输入尺寸 |  MOTA  |  IDF1  |  FPS   |  下载链接 | 配置文件 |
 | :-------------| :------- | :----: | :----: | :----: | :-----: |:------: |
-|  BDD100K      | 1088x608 |  34.9 |  39.9  |    -    | [下载链接](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_bdd100k_vehicle.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mot/vehicle/fairmot_dla34_30e_1088x608_bdd100k_vehicle.yml) |
-|  KITTI        | 1088x608 |  53.9 |    -   |    -   |[下载链接](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_kitti_vehicle.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mot/vehicle/fairmot_dla34_30e_1088x608_kitti_vehicle.yml) |
-|  VisDrone     | 1088x608 |  29.8 |   51.3 |    -   | [下载链接](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_visdrone_vehicle.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mot/vehicle/fairmot_dla34_30e_1088x608_visdrone_vehicle.yml) |
+|  BDD100K      | 1088x608 |  43.5 |  50.0  |    -    | [下载链接](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_bdd100kmot_vehicle.pdparams) | [配置文件](./vehicle/fairmot_dla34_30e_1088x608_bdd100kmot_vehicle.yml) |
+|  KITTI        | 1088x608 |  82.7 |    -   |    -   |[下载链接](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_kitti_vehicle.pdparams) | [配置文件](./vehicle/fairmot_dla34_30e_1088x608_kitti_vehicle.yml) |
+|  VisDrone     | 1088x608 |  52.1 |   63.3 |    -   | [下载链接](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_visdrone_vehicle.pdparams) | [配置文件](./vehicle/fairmot_dla34_30e_1088x608_visdrone_vehicle.yml) |
+
 
 
 ## 数据集准备

configs/mot/vehicle/README_cn.md

@@ -19,12 +19,12 @@
 
 |    数据集      |  输入尺寸 |  MOTA  |  IDF1  |  FPS   |  下载链接 | 配置文件 |
 | :-------------| :------- | :----: | :----: | :----: | :-----: |:------: |
-|  BDD100K      | 1088x608 |  34.9 |  39.9  |    -    | [下载链接](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_bdd100kmot_vehicle.pdparams) | [配置文件](./fairmot_dla34_30e_1088x608_bdd100kmot_vehicle.yml) |
-|  KITTI        | 1088x608 |  53.9 |    -   |    -   |[下载链接](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_kitti_vehicle.pdparams) | [配置文件](./fairmot_dla34_30e_1088x608_kitti_vehicle.yml) |
-|  VisDrone     | 1088x608 |  29.8 |   51.3 |    -   | [下载链接](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_visdrone_vehicle.pdparams) | [配置文件](./fairmot_dla34_30e_1088x608_visdrone_vehicle.yml) |
+|  BDD100K      | 1088x608 |  43.5 |  50.0  |    -    | [下载链接](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_bdd100kmot_vehicle.pdparams) | [配置文件](./fairmot_dla34_30e_1088x608_bdd100kmot_vehicle.yml) |
+|  KITTI        | 1088x608 |  82.7 |    -   |    -   |[下载链接](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_kitti_vehicle.pdparams) | [配置文件](./fairmot_dla34_30e_1088x608_kitti_vehicle.yml) |
+|  VisDrone     | 1088x608 |  52.1 |   63.3 |    -   | [下载链接](https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608_visdrone_vehicle.pdparams) | [配置文件](./fairmot_dla34_30e_1088x608_visdrone_vehicle.yml) |
 
 **注意:**
- FairMOT均使用DLA-34为骨干网络，2个GPU进行训练，每个GPU上batch size为6，训练30个epoch。
+ FairMOT均使用DLA-34为骨干网络，4个GPU进行训练，每个GPU上batch size为6，训练30个epoch。
 
 
 ## 数据集准备和处理

configs/mot/vehicle/fairmot_dla34_30e_1088x608_bdd100kmot_vehicle.yml

@@ -24,3 +24,17 @@ TestMOTDataset:
   !MOTImageFolder
     dataset_dir: dataset/mot
     keep_ori_im: True # set True if save visualization images or video
+
+# model config
+FairMOT:
+  detector: CenterNet
+  reid: FairMOTEmbeddingHead
+  loss: FairMOTLoss
+  tracker: JDETracker
+
+JDETracker:
+  min_box_area: 0
+  vertical_ratio: 0 # no need to filter bboxes according to w/h
+  conf_thres: 0.4
+  tracked_thresh: 0.4
+  metric_type: cosine

configs/mot/vehicle/fairmot_dla34_30e_1088x608_kitti_vehicle.yml

@@ -25,3 +25,17 @@ TestMOTDataset:
   !MOTImageFolder
     dataset_dir: dataset/mot
     keep_ori_im: True # set True if save visualization images or video
+
+# model config
+FairMOT:
+  detector: CenterNet
+  reid: FairMOTEmbeddingHead
+  loss: FairMOTLoss
+  tracker: JDETracker
+
+JDETracker:
+  min_box_area: 0
+  vertical_ratio: 0 # no need to filter bboxes according to w/h
+  conf_thres: 0.4
+  tracked_thresh: 0.4
+  metric_type: cosine

configs/mot/vehicle/fairmot_dla34_30e_1088x608_visdrone_vehicle.yml

@@ -24,3 +24,17 @@ TestMOTDataset:
   !MOTImageFolder
     dataset_dir: dataset/mot
     keep_ori_im: True # set True if save visualization images or video
+
+# model config
+FairMOT:
+  detector: CenterNet
+  reid: FairMOTEmbeddingHead
+  loss: FairMOTLoss
+  tracker: JDETracker
+
+JDETracker:
+  min_box_area: 0
+  vertical_ratio: 0 # no need to filter bboxes according to w/h
+  conf_thres: 0.4
+  tracked_thresh: 0.4
+  metric_type: cosine

deploy/python/mot_jde_infer.py

@@ -81,10 +81,14 @@ class JDE_Detector(Detector):
         assert batch_size == 1, "The JDE Detector only supports batch size=1 now"
         assert pred_config.tracker, "Tracking model should have tracker"
         tp = pred_config.tracker
+        min_box_area = tp['min_box_area'] if 'min_box_area' in tp else 200
+        vertical_ratio = tp['vertical_ratio'] if 'vertical_ratio' in tp else 1.6
         conf_thres = tp['conf_thres'] if 'conf_thres' in tp else 0.
         tracked_thresh = tp['tracked_thresh'] if 'tracked_thresh' in tp else 0.7
         metric_type = tp['metric_type'] if 'metric_type' in tp else 'euclidean'
         self.tracker = JDETracker(
+            min_box_area=min_box_area,
+            vertical_ratio=vertical_ratio,
             conf_thres=conf_thres,
             tracked_thresh=tracked_thresh,
             metric_type=metric_type)
@@ -102,11 +106,13 @@ class JDE_Detector(Detector):
             tid = t.track_id
             tscore = t.score
             if tscore < threshold: continue
-            vertical = tlwh[2] / tlwh[3] > 1.6
-            if tlwh[2] * tlwh[3] > self.tracker.min_box_area and not vertical:
-                online_tlwhs.append(tlwh)
-                online_ids.append(tid)
-                online_scores.append(tscore)
+            if tlwh[2] * tlwh[3] <= self.tracker.min_box_area: continue
+            if self.tracker.vertical_ratio > 0 and tlwh[2] / tlwh[
+                    3] > self.tracker.vertical_ratio:
+                continue
+            online_tlwhs.append(tlwh)
+            online_ids.append(tid)
+            online_scores.append(tscore)
         return online_tlwhs, online_scores, online_ids
 
     def predict(self, image_list, threshold=0.5, warmup=0, repeats=1):

deploy/python/tracker/jde_tracker.py

@@ -31,6 +31,9 @@ class JDETracker(object):
         det_thresh (float): threshold of detection score
         track_buffer (int): buffer for tracker
         min_box_area (int): min box area to filter out low quality boxes
+        vertical_ratio (float): w/h, the vertical ratio of the bbox to filter
+            bad results, set 1.6 default for pedestrian tracking. If set -1 
+            means no need to filter bboxes.
         tracked_thresh (float): linear assignment threshold of tracked 
             stracks and detections
         r_tracked_thresh (float): linear assignment threshold of 
@@ -47,6 +50,7 @@ class JDETracker(object):
                  det_thresh=0.3,
                  track_buffer=30,
                  min_box_area=200,
+                 vertical_ratio=1.6,
                  tracked_thresh=0.7,
                  r_tracked_thresh=0.5,
                  unconfirmed_thresh=0.7,
@@ -56,6 +60,8 @@ class JDETracker(object):
         self.det_thresh = det_thresh
         self.track_buffer = track_buffer
         self.min_box_area = min_box_area
+        self.vertical_ratio = vertical_ratio
+
         self.tracked_thresh = tracked_thresh
         self.r_tracked_thresh = r_tracked_thresh
         self.unconfirmed_thresh = unconfirmed_thresh

ppdet/engine/tracker.py

@@ -144,11 +144,13 @@ class Tracker(object):
                 tid = t.track_id
                 tscore = t.score
                 if tscore < draw_threshold: continue
-                vertical = tlwh[2] / tlwh[3] > 1.6
-                if tlwh[2] * tlwh[3] > tracker.min_box_area and not vertical:
-                    online_tlwhs.append(tlwh)
-                    online_ids.append(tid)
-                    online_scores.append(tscore)
+                if tlwh[2] * tlwh[3] <= tracker.min_box_area: continue
+                if tracker.vertical_ratio > 0 and tlwh[2] / tlwh[
+                        3] > tracker.vertical_ratio:
+                    continue
+                online_tlwhs.append(tlwh)
+                online_ids.append(tid)
+                online_scores.append(tscore)
             timer.toc()
 
             # save results

ppdet/metrics/mot_metrics.py

@@ -374,7 +374,7 @@ class KITTIEvaluation(object):
         # get number of sequences and
         # get number of frames per sequence from test mapping
         # (created while extracting the benchmark)
-        self.gt_path = os.path.join(gt_path, "label_02")
+        self.gt_path = os.path.join(gt_path, "../labels")
         self.n_frames = n_frames
         self.sequence_name = seqs
         self.n_sequences = n_sequences
@@ -1176,7 +1176,7 @@ class KITTIMOTMetric(Metric):
         assert data_type == 'kitti', "data_type should 'kitti'"
         self.result_root = result_root
         self.gt_path = data_root
-        gt_path = '{}/label_02/{}.txt'.format(data_root, seq)
+        gt_path = '{}/../labels/{}.txt'.format(data_root, seq)
         gt = open(gt_path, "r")
         max_frame = 0
         for line in gt:

ppdet/modeling/mot/tracker/jde_tracker.py

@@ -39,6 +39,9 @@ class JDETracker(object):
         det_thresh (float): threshold of detection score
         track_buffer (int): buffer for tracker
         min_box_area (int): min box area to filter out low quality boxes
+        vertical_ratio (float): w/h, the vertical ratio of the bbox to filter
+            bad results, set 1.6 default for pedestrian tracking. If set -1 
+            means no need to filter bboxes.
         tracked_thresh (float): linear assignment threshold of tracked 
             stracks and detections
         r_tracked_thresh (float): linear assignment threshold of 
@@ -55,6 +58,7 @@ class JDETracker(object):
                  det_thresh=0.3,
                  track_buffer=30,
                  min_box_area=200,
+                 vertical_ratio=1.6,
                  tracked_thresh=0.7,
                  r_tracked_thresh=0.5,
                  unconfirmed_thresh=0.7,
@@ -64,6 +68,8 @@ class JDETracker(object):
         self.det_thresh = det_thresh
         self.track_buffer = track_buffer
         self.min_box_area = min_box_area
+        self.vertical_ratio = vertical_ratio
+
         self.tracked_thresh = tracked_thresh
         self.r_tracked_thresh = r_tracked_thresh
         self.unconfirmed_thresh = unconfirmed_thresh