Localization-Guided Track: A Deep Association Multi-Object Tracking Framework Based on Localization Confidence of Detections

LG-Track is a simple, robust and reliable tracker

Localization-Guided Track: A Deep Association Multi-Object Tracking Framework Based on Localization Confidence of Detections

Update

[2023/12] we proposed Deep LG-Track, which makes LG-Track more robust. We will release the paper of Deep LG-Track as soon as we can.

Tracker	Dataset	HOTA	AssA	IDF1	MOTA	DetA
LG-Track	MOT17 test	65.4	65.4	80.4	81.4	65.6
Deep LG-Track	MOT17 test	65.9	66.5	81.4	81.3	65.5
LG-Track	MOT20 test	63.4	62.9	77.4	77.8	64.1
Deep LG-Track	MOT20 test	64.0	64.1	78.4	77.6	64.0

Abstract

In currently available literature, no tracking-by-detection (TBD) paradigm-based tracking method has considered the localization confidence of detection boxes. In most TBD-based methods, it is considered that objects of low detection confidence are highly occluded and thus it is a normal practice to directly disregard such objects or to reduce their priority in matching. In addition, appearance similarity is not a factor to consider for matching these objects. However, in terms of the detection confidence fusing classification and localization, objects of low detection confidence may have inaccurate localization but clear appearance; similarly, objects of high detection confidence may have inaccurate localization or unclear appearance; yet these objects are not further classified. In view of these issues, we propose Localization-Guided Track (LG-Track). Firstly, localization confidence is applied in MOT for the first time, with appearance clarity and localization accuracy of detection boxes taken into account, and an effective deep association mechanism is designed; secondly, based on the classification confidence and localization confidence, a more appropriate cost matrix can be selected and used; finally, extensive experiments have been conducted on MOT17 and MOT20 datasets. The results show that our proposed method outperforms the compared state-of-art tracking methods.

Tracking performance

Results on MOT challenge test set

Dataset	HOTA	MOTA	IDF1	AssA	DetA
MOT17	65.4	81.4	80.4	65.4	65.6
MOT20	63.4	77.8	77.4	62.9	64.1

Visualization results on MOT challenge test set

Installation

LG-Track code is based on BOT-SORT and ByteTrack. Visit their installation guides for more setup options.

Step 1. Install LG-Track.

git clone https://github.com/mengting2023/LG-Track.git
cd LG-Track
pip3 install -r requirements.txt
python3 setup.py develop

Step 2. Install pycocotools.

pip3 install cython; pip3 install 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

Step 3. Others

pip3 install cython_bbox
# faiss cpu / gpu
pip3 install faiss-cpu
pip3 install faiss-gpu

Data preparation

Download MOT17 and MOT20 from the official website. And put them put them under <LG-Track_dir>/datasets in the following structure:

<datasets>
      │
      ├── MOT17
      │      ├── train
      │      └── test    
      │
      └── MOT20
             ├── train
             └── test

Model and GMC_files

Download and store the trained models in 'weights' folder and GMC_files in 'tracker' floder as follow:

# models
<LG-Track_dir>/weights
# GMC_files
<LG-Track_dir>/tracker

We used the publicly available ByteTrack model zoo trained on MOT17 and MOT20 for YOLOX object detection.
We used the publicly available BOT-SORT model zoo trained on MOT17 and MOT20 for FastReID and GMC_files.

Tracking

Test on MOT17

cd <LG-Track_dir>
python3 main.py --datasets 'MOT17' --split 'test'
python3 tracker/tools/interpolation.py --txt_path <path_to_track_result>

Test on MOT20

cd <LG-Track_dir>
python3 main.py --datasets 'MOT20' --split 'test'
python3 tracker/tools/interpolation.py --txt_path <path_to_track_result>

if you have got the detection results (x1, y1, x2, y2, detection confidence, localization confidence, classification confidence) from other detectors, you can run:

cd <LG-Track_dir>
python3 track.py --datasets 'MOT20' --split 'test'
python3 tracker/tools/interpolation.py --txt_path <path_to_track_result>

Citation

if you think this work is useful, please consider to cite our paper:

@article{meng2023localization,
  title={Localization-Guided Track: A Deep Association Multi-Object Tracking Framework Based on Localization Confidence of Detections},
  author={Meng, Ting and Fu, Chunyun and Huang, Mingguang and Wang, Xiyang and He, Jiawei and Huang, Tao and Shi, Wankai},
  journal={arXiv preprint arXiv:2309.09765},
  year={2023}
}

Acknowledgement

The codebase is built highly upon BOT-SORT, ByteTrack, FastReID and YOLOX. Thanks for their excellent work!