Skip to content

Latest commit

 

History

History
208 lines (157 loc) · 13.9 KB

README.md

File metadata and controls

208 lines (157 loc) · 13.9 KB
Raw

English | 简体中文

FairMOT (FairMOT: On the Fairness of Detection and Re-Identification in Multiple Object Tracking)

Table of Contents

Introduction

FairMOT is based on an Anchor Free detector Centernet, which overcomes the problem of anchor and feature misalignment in anchor based detection framework. The fusion of deep and shallow features enables the detection and ReID tasks to obtain the required features respectively. It also uses low dimensional ReID features. FairMOT is a simple baseline composed of two homogeneous branches propose to predict the pixel level target score and ReID features. It achieves the fairness between the two tasks and obtains a higher level of real-time MOT performance.

PP-Tracking real-time MOT system

In addition, PaddleDetection also provides PP-Tracking real-time multi-object tracking system. PP-Tracking is the first open source real-time Multi-Object Tracking system, and it is based on PaddlePaddle deep learning framework. It has rich models, wide application and high efficiency deployment.

PP-Tracking supports two paradigms: single camera tracking (MOT) and multi-camera tracking (MTMCT). Aiming at the difficulties and pain points of actual business, PP-Tracking provides various MOT functions and applications such as pedestrian tracking, vehicle tracking, multi-class tracking, small object tracking, traffic statistics and multi-camera tracking. The deployment method supports API and GUI visual interface, and the deployment language supports Python and C++, The deployment platform environment supports Linux, NVIDIA Jetson, etc.

AI studio public project tutorial

PP-tracking provides an AI studio public project tutorial. Please refer to this tutorial.

Model Zoo

FairMOT Results on MOT-16 Training Set

backbone input shape MOTA IDF1 IDS FP FN FPS download config
DLA-34(paper) 1088x608 83.3 81.9 544 3822 14095 - - -
DLA-34 1088x608 83.2 83.1 499 3861 14223 - model config
DLA-34 864x480 80.8 81.1 561 3643 16967 - model config
DLA-34 576x320 74.0 76.1 640 4989 23034 - model config

FairMOT Results on MOT-16 Test Set

backbone input shape MOTA IDF1 IDS FP FN FPS download config
DLA-34(paper) 1088x608 74.9 72.8 1074 - - 25.9 - -
DLA-34 1088x608 75.0 74.7 919 7934 36747 - model config
DLA-34 864x480 73.0 72.6 977 7578 40601 - model config
DLA-34 576x320 69.9 70.2 1044 8869 44898 - model config

Notes:

  • FairMOT DLA-34 used 2 GPUs for training and mini-batch size as 6 on each GPU, and trained for 30 epoches.

FairMOT enhance model

Results on MOT-16 Test Set

backbone input shape MOTA IDF1 IDS FP FN FPS download config
DLA-34 1088x608 75.9 74.7 1021 11425 31475 - model config
HarDNet-85 1088x608 75.0 70.0 1050 11837 32774 - model config

Results on MOT-17 Test Set

backbone input shape MOTA IDF1 IDS FP FN FPS download config
DLA-34 1088x608 75.3 74.2 3270 29112 106749 - model config
HarDNet-85 1088x608 74.7 70.7 3210 29790 109914 - model config

Notes:

  • FairMOT enhance used 8 GPUs for training, and the crowdhuman dataset is added to the train-set during training.
  • For FairMOT enhance DLA-34 the batch size is 16 on each GPU，and trained for 60 epoches.
  • For FairMOT enhance HarDNet-85 the batch size is 10 on each GPU，and trained for 30 epoches.

FairMOT light model

Results on MOT-16 Test Set

backbone input shape MOTA IDF1 IDS FP FN FPS download config
HRNetV2-W18 1088x608 71.7 66.6 1340 8642 41592 - model config

Results on MOT-17 Test Set

backbone input shape MOTA IDF1 IDS FP FN FPS download config
HRNetV2-W18 1088x608 70.7 65.7 4281 22485 138468 - model config
HRNetV2-W18 864x480 70.3 65.8 4056 18927 144486 - model config
HRNetV2-W18 576x320 65.3 64.8 4137 28860 163017 - model config

Notes:

  • FairMOT HRNetV2-W18 used 8 GPUs for training and mini-batch size as 4 on each GPU, and trained for 30 epoches. Only ImageNet pre-train model is used, and the optimizer adopts Momentum. The crowdhuman dataset is added to the train-set during training.

FairMOT + BYTETracker

Results on MOT-17 Half Set

backbone input shape MOTA IDF1 IDS FP FN FPS download config
DLA-34 1088x608 69.1 72.8 299 1957 14412 - model config
DLA-34 + BYTETracker 1088x608 70.3 73.2 234 2176 13598 - model config

Notes:

  • FairMOT here is for ablation study, the training dataset is the 5 datasets of MIX(Caltech,CUHKSYSU,PRW,Cityscapes,ETHZ) and the first half of MOT17 Train, and the pretrain weights is CenterNet COCO model, the evaluation is on the second half of MOT17 Train.
  • BYTETracker adapt to other FairMOT models of PaddleDetection, you can modify the tracker of the config like this:
JDETracker:
 use_byte: True
 match_thres: 0.8
 conf_thres: 0.4
 low_conf_thres: 0.2

Fairmot transfer learning model

Results on GMOT-40 airplane subset

backbone input shape MOTA IDF1 IDS FP FN FPS download config
DLA-34 1088x608 96.6 94.7 19 300 466 - model config

Note:

  • The dataset of this model is a subset of airport category extracted from GMOT-40 dataset. The download link provided by the PaddleDetection team iswget https://bj.bcebos.com/v1/paddledet/data/mot/airplane.zip, unzip and store it in the dataset/mot, and then copy the airplane.train to dataset/mot/image_lists.
  • FairMOT model here uses the pedestrian FairMOT trained model for pre- training weights. The train-set used is the complete set of airplane, with a total of 4 video sequences, and it also used for evaluation.
  • When applied to the tracking other objects, you should modify min_box_area and vertical_ratio of the tracker in the corresponding config file, like this：
JDETracker:
 conf_thres: 0.4
 tracked_thresh: 0.4
 metric_type: cosine
 min_box_area: 0 # 200 for pedestrian
 vertical_ratio: 0 # 1.6 for pedestrian

Getting Start

1. Training

Training FairMOT on 2 GPUs with following command

python -m paddle.distributed.launch --log_dir=./fairmot_dla34_30e_1088x608/ --gpus 0,1 tools/train.py -c configs/mot/fairmot/fairmot_dla34_30e_1088x608.yml

2. Evaluation

Evaluating the track performance of FairMOT on val dataset in single GPU with following commands:

# use weights released in PaddleDetection model zoo
CUDA_VISIBLE_DEVICES=0 python tools/eval_mot.py -c configs/mot/fairmot/fairmot_dla34_30e_1088x608.yml -o weights=https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608.pdparams

# use saved checkpoint in training
CUDA_VISIBLE_DEVICES=0 python tools/eval_mot.py -c configs/mot/fairmot/fairmot_dla34_30e_1088x608.yml -o weights=output/fairmot_dla34_30e_1088x608/model_final.pdparams

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:
  !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

Inference a vidoe on single GPU with following command:

# inference on video and save a video
CUDA_VISIBLE_DEVICES=0 python tools/infer_mot.py -c configs/mot/fairmot/fairmot_dla34_30e_1088x608.yml -o weights=https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608.pdparams --video_file={your video name}.mp4  --save_videos

Notes:

  • Please make sure that ffmpeg is installed first, on Linux(Ubuntu) platform you can directly install it by the following command:apt-get update && apt-get install -y ffmpeg.

4. Export model

CUDA_VISIBLE_DEVICES=0 python tools/export_model.py -c configs/mot/fairmot/fairmot_dla34_30e_1088x608.yml -o weights=https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608.pdparams

5. Using exported model for python inference

python deploy/pptracking/python/mot_jde_infer.py --model_dir=output_inference/fairmot_dla34_30e_1088x608 --video_file={your video name}.mp4 --device=GPU --save_mot_txts

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

python deploy/python/mot_keypoint_unite_infer.py --mot_model_dir=output_inference/fairmot_dla34_30e_1088x608/ --keypoint_model_dir=output_inference/higherhrnet_hrnet_w32_512/ --video_file={your video name}.mp4 --device=GPU

Notes:

  • Keypoint model export tutorial: configs/keypoint/README.md.

Citations

@article{zhang2020fair,
  title={FairMOT: On the Fairness of Detection and Re-Identification in Multiple Object Tracking},
  author={Zhang, Yifu and Wang, Chunyu and Wang, Xinggang and Zeng, Wenjun and Liu, Wenyu},
  journal={arXiv preprint arXiv:2004.01888},
  year={2020}
}
@article{shao2018crowdhuman,
  title={CrowdHuman: A Benchmark for Detecting Human in a Crowd},
  author={Shao, Shuai and Zhao, Zijian and Li, Boxun and Xiao, Tete and Yu, Gang and Zhang, Xiangyu and Sun, Jian},
  journal={arXiv preprint arXiv:1805.00123},
  year={2018}
}