[MOT] Fix mot label list, add result txt format (#4682)

* fix label_list for mot, add warnings

* add result txt format, test=document_fix

* fix readme of mtmct, fix paddle.disable_static, test=document_fix

* remove paddle api

configs/mot/deepsort/README_cn.md

@@ -94,6 +94,7 @@ CUDA_VISIBLE_DEVICES=0 python tools/eval_mot.py -c configs/mot/deepsort/deepsort
  - JDE YOLOv3行人检测模型是和JDE和FairMOT使用同样的MOT数据集训练的，因此MOTA较高。而其他通用检测模型如PPYOLOv2只使用了MOT17 half数据集训练。
  - JDE YOLOv3模型与通用检测模型如YOLOv3和PPYOLOv2最大的区别是使用了JDEBBoxPostProcess后处理，结果输出坐标没有缩放回原图，而通用检测模型输出坐标是缩放回原图的。
  - `--scaled`表示在模型输出结果的坐标是否已经是缩放回原图的，如果使用的检测模型是JDE YOLOv3则为False，如果使用通用检测模型则为True, 默认值是False。
+ - 跟踪结果会存于`{output_dir}/mot_results/`中，里面每个视频序列对应一个txt，每个txt文件每行信息是`frame,id,x1,y1,w,h,score,-1,-1,-1`, 此外`{output_dir}`可通过`--output_dir`设置。
 
 ### 2. 预测
 
@@ -142,6 +143,7 @@ python deploy/python/mot_sde_infer.py --model_dir=output_inference/ppyolov2_r50v
 ```
 **注意:**
  跟踪模型是对视频进行预测，不支持单张图的预测，默认保存跟踪结果可视化后的视频，可添加`--save_mot_txts`(对每个视频保存一个txt)或`--save_mot_txt_per_img`(对每张图片保存一个txt)表示保存跟踪结果的txt文件，或`--save_images`表示保存跟踪结果可视化图片。
+ 跟踪结果txt文件每行信息是`frame,id,x1,y1,w,h,score,-1,-1,-1`。
  `--scaled`表示在模型输出结果的坐标是否已经是缩放回原图的，如果使用的检测模型是JDE的YOLOv3则为False，如果使用通用检测模型则为True。
 
 

configs/mot/fairmot/README.md

@@ -93,13 +93,14 @@ CUDA_VISIBLE_DEVICES=0 python tools/eval_mot.py -c configs/mot/fairmot/fairmot_d
 ```
 **Notes:**
   The default evaluation dataset is MOT-16 Train Set. If you want to change the evaluation dataset, please refer to the following code and modify `configs/datasets/mot.yml`：
-```
-EvalMOTDataset:
+  ```
+  EvalMOTDataset:
     !MOTImageFolder
       dataset_dir: dataset/mot
       data_root: MOT17/images/train
       keep_ori_im: False # set True if save visualization images or video
-```
+  ```
+  Tracking results will be saved in `{output_dir}/mot_results/`, and every sequence has one txt file, each line of the txt file is `frame,id,x1,y1,w,h,score,-1,-1,-1`, and you can set `{output_dir}` by `--output_dir`.
 
 ### 3. Inference
 
@@ -126,6 +127,7 @@ python deploy/python/mot_jde_infer.py --model_dir=output_inference/fairmot_dla34
 ```
 **Notes:**
 The tracking model is used to predict the video, and does not support the prediction of a single image. The visualization video of the tracking results is saved by default. You can add `--save_mot_txts` to save the txt result file, or `--save_images` to save the visualization images.
+Each line of the tracking results txt file is `frame,id,x1,y1,w,h,score,-1,-1,-1`.
 
 
 ### 6. Using exported MOT and keypoint model for unite python inference

configs/mot/fairmot/README_cn.md

@@ -91,13 +91,14 @@ CUDA_VISIBLE_DEVICES=0 python tools/eval_mot.py -c configs/mot/fairmot/fairmot_d
 ```
 **注意:**
  默认评估的是MOT-16 Train Set数据集, 如需换评估数据集可参照以下代码修改`configs/datasets/mot.yml`：
-```
-EvalMOTDataset:
+  ```
+  EvalMOTDataset:
     !MOTImageFolder
       dataset_dir: dataset/mot
       data_root: MOT17/images/train
       keep_ori_im: False # set True if save visualization images or video
-```
+  ```
+  跟踪结果会存于`{output_dir}/mot_results/`中，里面每个视频序列对应一个txt，每个txt文件每行信息是`frame,id,x1,y1,w,h,score,-1,-1,-1`, 此外`{output_dir}`可通过`--output_dir`设置。
 
 ### 3. 预测
 
@@ -124,6 +125,7 @@ python deploy/python/mot_jde_infer.py --model_dir=output_inference/fairmot_dla34
 ```
 **注意:**
  跟踪模型是对视频进行预测，不支持单张图的预测，默认保存跟踪结果可视化后的视频，可添加`--save_mot_txts`表示保存跟踪结果的txt文件，或`--save_images`表示保存跟踪结果可视化图片。
+ 跟踪结果txt文件每行信息是`frame,id,x1,y1,w,h,score,-1,-1,-1`。
 
 ### 6. 用导出的跟踪和关键点模型Python联合预测
 

configs/mot/headtracking21/README_cn.md

@@ -61,6 +61,7 @@ python deploy/python/mot_jde_infer.py --model_dir=output_inference/fairmot_dla34
 ```
 **注意:**
  跟踪模型是对视频进行预测，不支持单张图的预测，默认保存跟踪结果可视化后的视频，可添加`--save_mot_txts`表示保存跟踪结果的txt文件，或`--save_images`表示保存跟踪结果可视化图片。
+ 跟踪结果txt文件每行信息是`frame,id,x1,y1,w,h,score,-1,-1,-1`。
 
 ## 引用
 ```

configs/mot/jde/README.md

@@ -62,13 +62,14 @@ CUDA_VISIBLE_DEVICES=0 python tools/eval_mot.py -c configs/mot/jde/jde_darknet53
 ```
 **Notes:**
   The default evaluation dataset is MOT-16 Train Set. If you want to change the evaluation dataset, please refer to the following code and modify `configs/datasets/mot.yml`：
-```
-EvalMOTDataset:
+  ```
+  EvalMOTDataset:
     !MOTImageFolder
       dataset_dir: dataset/mot
       data_root: MOT17/images/train
       keep_ori_im: False # set True if save visualization images or video
-```
+  ```
+  Tracking results will be saved in `{output_dir}/mot_results/`, and every sequence has one txt file, each line of the txt file is `frame,id,x1,y1,w,h,score,-1,-1,-1`, and you can set `{output_dir}` by `--output_dir`.
 
 ### 3. Inference
 
@@ -95,6 +96,7 @@ python deploy/python/mot_jde_infer.py --model_dir=output_inference/jde_darknet53
 ```
 **Notes:**
 The tracking model is used to predict the video, and does not support the prediction of a single image. The visualization video of the tracking results is saved by default. You can add `--save_mot_txts` to save the txt result file, or `--save_images` to save the visualization images.
+Each line of the tracking results txt file is `frame,id,x1,y1,w,h,score,-1,-1,-1`.
 
 
 ## Citations

configs/mot/jde/README_cn.md

@@ -63,13 +63,14 @@ CUDA_VISIBLE_DEVICES=0 python tools/eval_mot.py -c configs/mot/jde/jde_darknet53
 ```
 **注意:**
   默认评估的是MOT-16 Train Set数据集, 如需换评估数据集可参照以下代码修改`configs/datasets/mot.yml`：
-```
-EvalMOTDataset:
+  ```
+  EvalMOTDataset:
     !MOTImageFolder
       dataset_dir: dataset/mot
       data_root: MOT17/images/train
       keep_ori_im: False # set True if save visualization images or video
 ```
+  跟踪结果会存于`{output_dir}/mot_results/`中，里面每个视频序列对应一个txt，每个txt文件每行信息是`frame,id,x1,y1,w,h,score,-1,-1,-1`, 此外`{output_dir}`可通过`--output_dir`设置。
 
 ### 3. 预测
 
@@ -95,7 +96,9 @@ CUDA_VISIBLE_DEVICES=0 python tools/export_model.py -c configs/mot/jde/jde_darkn
 python deploy/python/mot_jde_infer.py --model_dir=output_inference/jde_darknet53_30e_1088x608 --video_file={your video name}.mp4 --device=GPU --save_mot_txts
 ```
 **注意:**
- 跟踪模型是对视频进行预测，不支持单张图的预测，默认保存跟踪结果可视化后的视频，可添加`--save_mot_txts`表示保存跟踪结果的txt文件，或`--save_images`表示保存跟踪结果可视化图片。。
+ 跟踪模型是对视频进行预测，不支持单张图的预测，默认保存跟踪结果可视化后的视频，可添加`--save_mot_txts`表示保存跟踪结果的txt文件，或`--save_images`表示保存跟踪结果可视化图片。
+ 跟踪结果txt文件每行信息是`frame,id,x1,y1,w,h,score,-1,-1,-1`。
+
 
 ## 引用
 ```

configs/mot/mcfairmot/README.md

@@ -44,13 +44,14 @@ CUDA_VISIBLE_DEVICES=0 python tools/eval_mot.py -c configs/mot/mcfairmot/mcfairm
 ```
 **Notes:**
   The default evaluation dataset is VisDrone2019 MOT val-set. If you want to change the evaluation dataset, please refer to the following code and modify `configs/datasets/mcmot.yml`：
-```
-EvalMOTDataset:
+  ```
+  EvalMOTDataset:
     !MOTImageFolder
       dataset_dir: dataset/mot
       data_root: your_dataset/images/val
       keep_ori_im: False # set True if save visualization images or video
-```
+  ```
+  Tracking results will be saved in `{output_dir}/mot_results/`, and every sequence has one txt file, each line of the txt file is `frame,id,x1,y1,w,h,score,cls_id,-1,-1`, and you can set `{output_dir}` by `--output_dir`.
 
 ### 3. Inference
 Inference a vidoe on single GPU with following command:
@@ -73,6 +74,7 @@ python deploy/python/mot_jde_infer.py --model_dir=output_inference/mcfairmot_dla
 ```
 **Notes:**
 The tracking model is used to predict the video, and does not support the prediction of a single image. The visualization video of the tracking results is saved by default. You can add `--save_mot_txts` to save the txt result file, or `--save_images` to save the visualization images.
+Each line of the tracking results txt file is `frame,id,x1,y1,w,h,score,cls_id,-1,-1`.
 
 
 ## Citations

configs/mot/mcfairmot/README_cn.md

@@ -44,13 +44,14 @@ CUDA_VISIBLE_DEVICES=0 python tools/eval_mot.py -c configs/mot/mcfairmot/mcfairm
 ```
 **注意:**
   默认评估的是VisDrone2019 MOT val-set数据集, 如需换评估数据集可参照以下代码修改`configs/datasets/mcmot.yml`：
-```
-EvalMOTDataset:
+  ```
+  EvalMOTDataset:
     !MOTImageFolder
       dataset_dir: dataset/mot
       data_root: your_dataset/images/val
       keep_ori_im: False # set True if save visualization images or video
-```
+  ```
+  多类别跟踪结果会存于`{output_dir}/mot_results/`中，里面每个视频序列对应一个txt，每个txt文件每行信息是`frame,id,x1,y1,w,h,score,cls_id,-1,-1`, 此外`{output_dir}`可通过`--output_dir`设置。
 
 ### 3. 预测
 使用单个GPU通过如下命令预测一个视频，并保存为视频
@@ -72,6 +73,7 @@ python deploy/python/mot_jde_infer.py --model_dir=output_inference/mcfairmot_dla
 ```
 **注意:**
  跟踪模型是对视频进行预测，不支持单张图的预测，默认保存跟踪结果可视化后的视频，可添加`--save_mot_txts`表示保存跟踪结果的txt文件，或`--save_images`表示保存跟踪结果可视化图片。
+ 多类别跟踪结果txt文件每行信息是`frame,id,x1,y1,w,h,score,cls_id,-1,-1`。
 
 
 ## 引用

configs/mot/mtmct/README_cn.md

@@ -90,13 +90,14 @@ tar -xvf deepsort_pplcnet_vehicle.tar
 ### 2. 用导出的模型基于Python去预测
 ```bash
 # 用导出PicoDet车辆检测模型和PPLCNet车辆ReID模型
-python deploy/pptracking/python/mot_sde_infer.py --model_dir=picodet_l_640_aic21mtmct_vehicle/ --reid_model_dir=deepsort_pplcnet_vehicle/ --mtmct_dir={your mtmct scene video folder} --device=GPU --scaled=True --save_mot_txts --save_images
+python deploy/pptracking/python/mot_sde_infer.py --model_dir=picodet_l_640_aic21mtmct_vehicle/ --reid_model_dir=deepsort_pplcnet_vehicle/ --mtmct_dir={your mtmct scene video folder} --mtmct_cfg=mtmct_cfg --device=GPU --scaled=True --save_mot_txts --save_images
 ```
 **注意:**
   跟踪模型是对视频进行预测，不支持单张图的预测，默认保存跟踪结果可视化后的视频，可添加`--save_mot_txts`(对每个视频保存一个txt)，或`--save_images`表示保存跟踪结果可视化图片。
   `--scaled`表示在模型输出结果的坐标是否已经是缩放回原图的，如果使用的检测模型是JDE的YOLOv3则为False，如果使用通用检测模型则为True。
   `--mtmct_dir`是MTMCT预测的某个场景的文件夹名字，里面包含该场景不同摄像头拍摄视频的图片文件夹，其数量至少为两个。
-  MTMCT跨镜头跟踪输出结果为视频和txt形式。每个图片文件夹各生成一个可视化的跨镜头跟踪结果，与单镜头跟踪的结果是不同的，单镜头跟踪的结果在几个视频文件夹间是独立无关的。MTMCT的结果txt只有一个，比单镜头跟踪结果txt多了第一列镜头id号。
+  `--mtmct_cfg`是MTMCT预测的某个场景的配置文件，里面包含该一些trick操作的开关和该场景摄像头相关设置的文件路径，用户可以自行更改相关路径以及设置某些操作是否启用。
+  MTMCT跨镜头跟踪输出结果为视频和txt形式。每个图片文件夹各生成一个可视化的跨镜头跟踪结果，与单镜头跟踪的结果是不同的，单镜头跟踪的结果在几个视频文件夹间是独立无关的。MTMCT的结果txt只有一个，比单镜头跟踪结果txt多了第一列镜头id号，跨镜头跟踪结果txt文件每行信息是`carame_id,frame,id,x1,y1,w,h,-1,-1`。
   MTMCT是[PP-Tracking](../../../deploy/pptracking)项目中的一个非常重要的方向，具体可前往该目录使用。
 
 

configs/mot/pedestrian/README_cn.md

@@ -99,6 +99,7 @@ python deploy/python/mot_jde_infer.py --model_dir=output_inference/fairmot_dla34
 ```
 **注意:**
  跟踪模型是对视频进行预测，不支持单张图的预测，默认保存跟踪结果可视化后的视频，可添加`--save_mot_txts`表示保存跟踪结果的txt文件，或`--save_images`表示保存跟踪结果可视化图片。
+ 跟踪结果txt文件每行信息是`frame,id,x1,y1,w,h,score,-1,-1,-1`。
 
 ## 引用
 ```

configs/mot/vehicle/README_cn.md

@@ -131,6 +131,7 @@ python deploy/python/mot_jde_infer.py --model_dir=output_inference/fairmot_dla34
 ```
 **注意:**
  跟踪模型是对视频进行预测，不支持单张图的预测，默认保存跟踪结果可视化后的视频，可添加`--save_mot_txts`表示保存跟踪结果的txt文件，或`--save_images`表示保存跟踪结果可视化图片。
+ 跟踪结果txt文件每行信息是`frame,id,x1,y1,w,h,score,-1,-1,-1`。
 
 ## 引用
 ```

deploy/pptracking/python/README.md

@@ -23,6 +23,7 @@ python deploy/pptracking/python/mot_jde_infer.py --model_dir=output_inference/fa
 ```
 **注意:**
  - 跟踪模型是对视频进行预测，不支持单张图的预测，默认保存跟踪结果可视化后的视频，可添加`--save_mot_txts`表示保存跟踪结果的txt文件，或`--save_images`表示保存跟踪结果可视化图片。
+ - 跟踪结果txt文件每行信息是`frame,id,x1,y1,w,h,score,-1,-1,-1`。
  - 对于多类别或车辆的FairMOT模型的导出和Python预测只需更改相应的config和模型权重即可。如：
  ```
  job_name=mcfairmot_hrnetv2_w18_dlafpn_30e_576x320_visdrone
@@ -32,6 +33,7 @@ python deploy/pptracking/python/mot_jde_infer.py --model_dir=output_inference/fa
  CUDA_VISIBLE_DEVICES=0 python tools/export_model.py -c ${config} -o weights=https://paddledet.bj.bcebos.com/models/mot/${job_name}.pdparams
  python deploy/pptracking/python/mot_jde_infer.py --model_dir=output_inference/${job_name} --video_file={your video name}.mp4 --device=GPU --save_mot_txts
  ```
+ - 多类别跟踪结果txt文件每行信息是`frame,id,x1,y1,w,h,score,cls_id,-1,-1`。
 
 
 ## 2. 对DeepSORT模型的导出和预测
@@ -65,6 +67,7 @@ python deploy/pptracking/python/mot_sde_infer.py --model_dir=output_inference/pp
 ```
 **注意:**
  - 跟踪模型是对视频进行预测，不支持单张图的预测，默认保存跟踪结果可视化后的视频，可添加`--save_mot_txts`(对每个视频保存一个txt)或`--save_images`表示保存跟踪结果可视化图片。
+ - 跟踪结果txt文件每行信息是`frame,id,x1,y1,w,h,score,-1,-1,-1`。
  - `--scaled`表示在模型输出结果的坐标是否已经是缩放回原图的，如果使用的检测模型是JDE的YOLOv3则为False，如果使用通用检测模型则为True。
 
 
@@ -88,8 +91,10 @@ python deploy/pptracking/python/mot_sde_infer.py --model_dir=picodet_l_640_aic21
 ```
 **注意:**
   跟踪模型是对视频进行预测，不支持单张图的预测，默认保存跟踪结果可视化后的视频，可添加`--save_mot_txts`(对每个视频保存一个txt)，或`--save_images`表示保存跟踪结果可视化图片。
+  跨镜头跟踪结果txt文件每行信息是`carame_id,frame,id,x1,y1,w,h,-1,-1`。
   `--scaled`表示在模型输出结果的坐标是否已经是缩放回原图的，如果使用的检测模型是JDE的YOLOv3则为False，如果使用通用检测模型则为True。
   `--mtmct_dir`是MTMCT预测的某个场景的文件夹名字，里面包含该场景不同摄像头拍摄视频的图片文件夹，其数量至少为两个。
+  `--mtmct_cfg`是MTMCT预测的某个场景的配置文件，里面包含该一些trick操作的开关和该场景摄像头相关设置的文件路径，用户可以自行更改相关路径以及设置某些操作是否启用。
 
 
 ## 参数说明:

deploy/pptracking/python/mot/matching/jde_matching.py

@@ -20,6 +20,8 @@ import scipy
 import numpy as np
 from scipy.spatial.distance import cdist
 from ..motion import kalman_filter
+import warnings
+warnings.filterwarnings("ignore")
 
 __all__ = [
     'merge_matches',

deploy/pptracking/python/mot/mtmct/camera_utils.py

@@ -12,7 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-https://github.com/LCFractal/AIC21-MTMC/tree/main/reid/reid-matching/tools
+This code is based on https://github.com/LCFractal/AIC21-MTMC/tree/main/reid/reid-matching/tools
+
+Note: The following codes are strongly related to carame parameters of the AIC21 test-set S06,
+    so they can only be used in S06, and can not be used for other MTMCT datasets.
 """
 
 import numpy as np

deploy/pptracking/python/mot/mtmct/postprocess.py

@@ -12,16 +12,13 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-https://github.com/LCFractal/AIC21-MTMC/tree/main/reid/reid-matching/tools
+This code is based on https://github.com/LCFractal/AIC21-MTMC/tree/main/reid/reid-matching/tools
 """
 
+import os
 import re
 import cv2
 from tqdm import tqdm
-import pickle
-import os
-import os.path as osp
-from os.path import join as opj
 import numpy as np
 import motmetrics as mm
 from functools import reduce
@@ -216,9 +213,6 @@ def get_mtmct_matching_results(pred_mtmct_file, secs_interval=0.5,
     res = np.loadtxt(pred_mtmct_file)  # 'cid, tid, fid, x1, y1, w, h, -1, -1'
     carame_ids = list(map(int, np.unique(res[:, 0])))
 
-    num_track_ids = int(np.max(res[:, 1]))
-    num_frames = int(np.max(res[:, 2]))
-
     res = res[:, :7]
     # each line in res: 'cid, tid, fid, x1, y1, w, h'
 
@@ -236,7 +230,7 @@ def get_mtmct_matching_results(pred_mtmct_file, secs_interval=0.5,
         print(
             'No common tracked ids in these videos, please check your MOT result or select new videos.'
         )
-        return None
+        return None, None
 
     # get mtmct matching results by cid_tid_fid_results[c_id][t_id][f_id]
     cid_tid_fid_results = dict()
@@ -302,8 +296,9 @@ def save_mtmct_crops(cid_tid_fid_res,
                 
                 im = cv2.imread(im_path)  # (H, W, 3)
                 
-                track = cid_tid_fid_res[c_id][t_id][f_id][
-                    0]  # only select one track
+                # only select one track
+                track = cid_tid_fid_res[c_id][t_id][f_id][0]
+
                 cid, tid, fid, x1, y1, w, h = [int(v) for v in track]
                 clip = im[y1:(y1 + h), x1:(x1 + w)]
                 clip = cv2.resize(clip, (width, height))

deploy/pptracking/python/mot/mtmct/utils.py

@@ -12,17 +12,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-https://github.com/LCFractal/AIC21-MTMC/tree/main/reid/reid-matching/tools
+This code is based on https://github.com/LCFractal/AIC21-MTMC/tree/main/reid/reid-matching/tools
 """
 
 import os
 import re
 import cv2
-import paddle
+import gc
 import numpy as np
 from sklearn import preprocessing
 from sklearn.cluster import AgglomerativeClustering
-import gc
 import motmetrics as mm
 import pandas as pd
 from tqdm import tqdm
@@ -183,14 +182,40 @@ def run_fac(prb_feats,
 def euclidean_distance(qf, gf):
     m = qf.shape[0]
     n = gf.shape[0]
-    dist_mat = 2 - 2 * paddle.matmul(qf, gf.t())
+    dist_mat = 2 - 2 * np.matmul(qf, gf.T)
     return dist_mat
 
 
-def batch_paddle_topk(qf, gf, k1, N=6000):
+def find_topk(a, k, axis=-1, largest=True, sorted=True):
+    if axis is None:
+        axis_size = a.size
+    else:
+        axis_size = a.shape[axis]
+    assert 1 <= k <= axis_size
+
+    a = np.asanyarray(a)
+    if largest:
+        index_array = np.argpartition(a, axis_size-k, axis=axis)
+        topk_indices = np.take(index_array, -np.arange(k)-1, axis=axis)
+    else:
+        index_array = np.argpartition(a, k-1, axis=axis)
+        topk_indices = np.take(index_array, np.arange(k), axis=axis)
+    topk_values = np.take_along_axis(a, topk_indices, axis=axis)
+    if sorted:
+        sorted_indices_in_topk = np.argsort(topk_values, axis=axis)
+        if largest:
+            sorted_indices_in_topk = np.flip(sorted_indices_in_topk, axis=axis)
+        sorted_topk_values = np.take_along_axis(
+            topk_values, sorted_indices_in_topk, axis=axis)
+        sorted_topk_indices = np.take_along_axis(
+            topk_indices, sorted_indices_in_topk, axis=axis)
+        return sorted_topk_values, sorted_topk_indices
+    return topk_values, topk_indices
+
+
+def batch_numpy_topk(qf, gf, k1, N=6000):
     m = qf.shape[0]
     n = gf.shape[0]
-    dist_mat = []
     initial_rank = []
     for j in range(n // N + 1):
         temp_gf = gf[j * N:j * N + N]
@@ -199,17 +224,16 @@ def batch_paddle_topk(qf, gf, k1, N=6000):
             temp_qf = qf[i * N:i * N + N]
             temp_d = euclidean_distance(temp_qf, temp_gf)
             temp_qd.append(temp_d)
-        temp_qd = paddle.concat(temp_qd, axis=0)
-        temp_qd = temp_qd / (paddle.max(temp_qd, axis=0)[0])
-        temp_qd = temp_qd.t()
+        temp_qd = np.concatenate(temp_qd, axis=0)
+        temp_qd = temp_qd / (np.max(temp_qd, axis=0)[0])
+        temp_qd = temp_qd.T
         initial_rank.append(
-            paddle.topk(
-                temp_qd, k=k1, axis=1, largest=False, sorted=True)[1])
+            find_topk(temp_qd, k=k1, axis=1, largest=False, sorted=True)[1])
     del temp_qd
     del temp_gf
     del temp_qf
     del temp_d
-    initial_rank = paddle.concat(initial_rank, axis=0).cpu().numpy()
+    initial_rank = np.concatenate(initial_rank, axis=0)
     return initial_rank
 
 
@@ -224,14 +248,14 @@ def batch_euclidean_distance(qf, gf, N=6000):
             temp_qf = qf[i * N:i * N + N]
             temp_d = euclidean_distance(temp_qf, temp_gf)
             temp_qd.append(temp_d)
-        temp_qd = paddle.concat(temp_qd, axis=0)
-        temp_qd = temp_qd / (paddle.max(temp_qd, axis=0)[0])
-        dist_mat.append(temp_qd.t())  # transpose
+        temp_qd = np.concatenate(temp_qd, axis=0)
+        temp_qd = temp_qd / (np.max(temp_qd, axis=0)[0])
+        dist_mat.append(temp_qd.T)
     del temp_qd
     del temp_gf
     del temp_qf
     del temp_d
-    dist_mat = paddle.concat(dist_mat, axis=0)
+    dist_mat = np.concatenate(dist_mat, axis=0)
     return dist_mat
 
 
@@ -239,14 +263,13 @@ def batch_v(feat, R, all_num):
     V = np.zeros((all_num, all_num), dtype=np.float32)
     m = feat.shape[0]
     for i in tqdm(range(m)):
-        temp_gf = feat[i].unsqueeze(0)
+        temp_gf = feat[i].reshape(1, -1)
         temp_qd = euclidean_distance(temp_gf, feat)
-        temp_qd = temp_qd / (paddle.max(temp_qd))
-        temp_qd = temp_qd.squeeze()
-        temp_qd = temp_qd.numpy()[R[i].tolist()]
-        temp_qd = paddle.to_tensor(temp_qd)
-        weight = paddle.exp(-temp_qd)
-        weight = (weight / paddle.sum(weight)).numpy()
+        temp_qd = temp_qd / (np.max(temp_qd))
+        temp_qd = temp_qd.reshape(-1)
+        temp_qd = temp_qd[R[i].tolist()]
+        weight = np.exp(-temp_qd)
+        weight = weight / np.sum(weight)
         V[i, R[i]] = weight.astype(np.float32)
     return V
 
@@ -259,13 +282,11 @@ def k_reciprocal_neigh(initial_rank, i, k1):
 
 
 def ReRank2(probFea, galFea, k1=20, k2=6, lambda_value=0.3):
-    # The following naming, e.g. gallery_num, is different from outer scope.
-    # Don't care about it.
     query_num = probFea.shape[0]
     all_num = query_num + galFea.shape[0]
-    feat = paddle.concat([probFea, galFea], axis=0)
-    initial_rank = batch_paddle_topk(feat, feat, k1 + 1, N=6000)
-    # del feat
+    feat = np.concatenate((probFea, galFea), axis=0)
+
+    initial_rank = batch_numpy_topk(feat, feat, k1 + 1, N=6000)
     del probFea
     del galFea
     gc.collect()  # empty memory
@@ -292,7 +313,8 @@ def ReRank2(probFea, galFea, k1=20, k2=6, lambda_value=0.3):
     del R
     gc.collect()  # empty memory
     initial_rank = initial_rank[:, :k2]
-    ### Faster version
+
+    # Faster version
     if k2 != 1:
         V_qe = np.zeros_like(V, dtype=np.float16)
         for i in range(all_num):
@@ -315,7 +337,7 @@ def ReRank2(probFea, galFea, k1=20, k2=6, lambda_value=0.3):
         jaccard_dist[i] = 1 - temp_min / (2. - temp_min)
     del V
     gc.collect()  # empty memory
-    original_dist = batch_euclidean_distance(feat, feat[:query_num, :]).numpy()
+    original_dist = batch_euclidean_distance(feat, feat[:query_num, :])
     final_dist = jaccard_dist * (1 - lambda_value
                                  ) + original_dist * lambda_value
     del original_dist
@@ -341,21 +363,16 @@ def visual_rerank(prb_feats,
         prb_feats, gal_feats = run_fac(prb_feats, gal_feats, prb_labels,
                                        gal_labels, 0.08, 20, 0.5, 1, 1)
     if use_rerank:
-        paddle.enable_static()
         print('current use rerank finetuned parameters....')
         # Step2: k-reciprocal. finetuned parameters: [k1,k2,lambda_value]
-        sims = ReRank2(
-            paddle.to_tensor(prb_feats),
-            paddle.to_tensor(gal_feats), 20, 3, 0.3)
+        sims = ReRank2(prb_feats, gal_feats, 20, 3, 0.3)
     else:
-        # sims = ComputeEuclid(prb_feats, gal_feats, 1)
         sims = 1.0 - np.dot(prb_feats, gal_feats.T)
 
     # NOTE: sims here is actually dist, the smaller the more similar
     return 1.0 - sims
 
 
-# sub_cluster
 def normalize(nparray, axis=0):
     nparray = preprocessing.normalize(nparray, norm='l2', axis=axis)
     return nparray
@@ -512,10 +529,6 @@ def get_labels(cid_tid_dict,
                use_rerank=True,
                use_st_filter=False):
     # 1st cluster
-    sub_cid_tids = list(cid_tid_dict.keys())
-    sub_labels = dict()
-    dis_thrs = [0.7, 0.5, 0.5, 0.5, 0.5, 0.7, 0.5, 0.5, 0.5, 0.5]
-
     sim_matrix = get_sim_matrix(
         cid_tid_dict,
         cid_tids,
@@ -532,7 +545,6 @@ def get_labels(cid_tid_dict,
 
     # 2nd cluster
     cid_tid_dict_new = combin_feature(cid_tid_dict, sub_cluster)
-    sub_labels = dict()
     sim_matrix = get_sim_matrix(
         cid_tid_dict_new,
         cid_tids,

deploy/pptracking/python/mot/mtmct/zone.py

@@ -12,7 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-https://github.com/LCFractal/AIC21-MTMC/tree/main/reid/reid-matching/tools
+This code is based on https://github.com/LCFractal/AIC21-MTMC/tree/main/reid/reid-matching/tools
+
+Note: The following codes are strongly related to zone of the AIC21 test-set S06,
+    so they can only be used in S06, and can not be used for other MTMCT datasets.
 """
 
 import os

deploy/pptracking/python/mtmct_cfg.yml

 # config for MTMCT
 MTMCT: True
-cameras_bias:
-  c041: 0
-  c042: 0
+cameras_bias: # default for scene S01. For S06, should modify as 'c041: 0  c042: 0'
+  c003: 0
+  c004: 0
 # 1.zone releated parameters
-use_zone: True
+use_zone: False
 zone_path: dataset/mot/aic21mtmct_vehicle/S06/zone
 # 2.tricks parameters, can be used for other mtmct dataset
 use_ff: True
 use_rerank: True
 # 3.camera releated parameters
-use_camera: True
+use_camera: False
 use_st_filter: False
 # 4.zone releated parameters
-use_roi: True
+use_roi: False
 roi_dir: dataset/mot/aic21mtmct_vehicle/S06

ppdet/data/source/category.py

@@ -92,19 +92,43 @@ def get_categories(metric_type, anno_file=None, arch=None):
         return (None, {'id': 'keypoint'})
 
     elif metric_type.lower() in ['mot', 'motdet', 'reid']:
-        return _mot_category()
+        if anno_file and os.path.isfile(anno_file):
+            cats = []
+            with open(anno_file) as f:
+                for line in f.readlines():
+                    cats.append(line.strip())
+            if cats[0] == 'background':
+                cats = cats[1:]
+            clsid2catid = {i: i for i in range(len(cats))}
+            catid2name = {i: name for i, name in enumerate(cats)}
+            return clsid2catid, catid2name
+        # anno file not exist, load default category 'pedestrian'.
+        else:
+            return _mot_category(category='pedestrian')
 
     elif metric_type.lower() in ['kitti', 'bdd100kmot']:
-        return _mot_category(category='car')
+        return _mot_category(category='vehicle')
 
     elif metric_type.lower() in ['mcmot']:
+        if anno_file and os.path.isfile(anno_file):
+            cats = []
+            with open(anno_file) as f:
+                for line in f.readlines():
+                    cats.append(line.strip())
+            if cats[0] == 'background':
+                cats = cats[1:]
+            clsid2catid = {i: i for i in range(len(cats))}
+            catid2name = {i: name for i, name in enumerate(cats)}
+            return clsid2catid, catid2name
+        # anno file not exist, load default categories of visdrone all 10 categories
+        else:
             return _visdrone_category()
 
     else:
         raise ValueError("unknown metric type {}".format(metric_type))
 
 
-def _mot_category(category='person'):
+def _mot_category(category='pedestrian'):
     """
     Get class id to category id map and category id
     to category name map of mot dataset

ppdet/data/source/mot.py

@@ -93,8 +93,10 @@ class MOTDataSet(DetDataset):
         if self.image_lists == []:
             return
         # only used to get categories and metric
-        return os.path.join(self.dataset_dir, 'image_lists',
-                            self.image_lists[0])
+        # only check first data, but the label_list of all data should be same.
+        first_mot_data = self.image_lists[0].split('.')[0]
+        anno_file = os.path.join(self.dataset_dir, first_mot_data, 'label_list.txt')
+        return anno_file
 
     def parse_dataset(self):
         self.img_files = OrderedDict()
@@ -272,8 +274,10 @@ class MCMOTDataSet(DetDataset):
         if self.image_lists == []:
             return
         # only used to get categories and metric
-        return os.path.join(self.dataset_dir, 'image_lists',
-                            self.image_lists[0])
+        # only check first data, but the label_list of all data should be same.
+        first_mot_data = self.image_lists[0].split('.')[0]
+        anno_file = os.path.join(self.dataset_dir, first_mot_data, 'label_list.txt')
+        return anno_file
 
     def parse_dataset(self):
         self.img_files = OrderedDict()

ppdet/modeling/mot/matching/deepsort_matching.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-This code is borrow from https://github.com/nwojke/deep_sort/tree/master/deep_sort
+This code is based on https://github.com/nwojke/deep_sort/tree/master/deep_sort
 """
 
 import numpy as np

ppdet/modeling/mot/matching/jde_matching.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-This code is borrow from https://github.com/Zhongdao/Towards-Realtime-MOT/blob/master/tracker/matching.py
+This code is based on https://github.com/Zhongdao/Towards-Realtime-MOT/blob/master/tracker/matching.py
 """
 
 import lap
@@ -20,9 +20,8 @@ import scipy
 import numpy as np
 from scipy.spatial.distance import cdist
 from ..motion import kalman_filter
-
-from ppdet.utils.logger import setup_logger
-logger = setup_logger(__name__)
+import warnings
+warnings.filterwarnings("ignore")
 
 __all__ = [
     'merge_matches',
@@ -76,7 +75,7 @@ def cython_bbox_ious(atlbrs, btlbrs):
     try:
         import cython_bbox
     except Exception as e:
-        logger.error('cython_bbox not found, please install cython_bbox.'
+        print('cython_bbox not found, please install cython_bbox.'
                      'for example: `pip install cython_bbox`.')
         raise e
 

ppdet/modeling/mot/motion/kalman_filter.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-This code is borrow from https://github.com/nwojke/deep_sort/blob/master/deep_sort/kalman_filter.py
+This code is based on https://github.com/nwojke/deep_sort/blob/master/deep_sort/kalman_filter.py
 """
 
 import numpy as np

ppdet/modeling/mot/tracker/base_jde_tracker.py

@@ -11,12 +11,17 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+"""
+This code is based on https://github.com/Zhongdao/Towards-Realtime-MOT/blob/master/tracker/multitracker.py
+"""
 
 import numpy as np
 from collections import defaultdict
 from collections import deque, OrderedDict
 from ..matching import jde_matching as matching
 from ppdet.core.workspace import register, serializable
+import warnings
+warnings.filterwarnings("ignore")
 
 __all__ = [
     'TrackState',

ppdet/modeling/mot/tracker/base_sde_tracker.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-This code is borrow from https://github.com/nwojke/deep_sort/blob/master/deep_sort/track.py
+This code is based on https://github.com/nwojke/deep_sort/blob/master/deep_sort/track.py
 """
 
 import datetime

ppdet/modeling/mot/tracker/deepsort_tracker.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """
-This code is borrow from https://github.com/nwojke/deep_sort/blob/master/deep_sort/tracker.py
+This code is based on https://github.com/nwojke/deep_sort/blob/master/deep_sort/tracker.py
 """
 
 import numpy as np

ppdet/modeling/mot/tracker/jde_tracker.py

@@ -17,7 +17,6 @@ This code is based on https://github.com/Zhongdao/Towards-Realtime-MOT/blob/mast
 
 import numpy as np
 from collections import defaultdict
-import paddle
 
 from ..matching import jde_matching as matching
 from ..motion import KalmanFilter