Skip to content

BoT-SORT: Robust Associations Multi-Pedestrian Tracking

License

Notifications You must be signed in to change notification settings

DenisKuplyakov/BoT-SORT

Β 
Β 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

Β 

History

22 Commits
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation

BoT-SORT

BoT-SORT: Robust Associations Multi-Pedestrian Tracking

Nir Aharon, Roy Orfaig, Ben-Zion Bobrovsky

PWC

PWC

https://arxiv.org/abs/2206.14651

Highlights πŸš€

  • YOLOX & YOLOv7 support
  • Multi-class support
  • Camera motion compensation
  • Re-identification

Coming Soon

  • Trained YOLOv7 models for MOTChallenge.
  • YOLOv7 detector.
  • Multi-class support.
  • Create OpenCV VideoStab GMC python binding or write Python version.
  • Deployment code.

Abstract

The goal of multi-object tracking (MOT) is detecting and tracking all the objects in a scene, while keeping a unique identifier for each object. In this paper, we present a new robust state-of-the-art tracker, which can combine the advantages of motion and appearance information, along with camera-motion compensation, and a more accurate Kalman filter state vector. Our new trackers BoT-SORT, and BoT-SORT-ReID rank first in the datasets of MOTChallenge [29, 11] on both MOT17 and MOT20 test sets, in terms of all the main MOT metrics: MOTA, IDF1, and HOTA. For MOT17: 80.5 MOTA, 80.2 IDF1, and 65.0 HOTA are achieved.

Visualization results on MOT challenge test set

MOT20-06.mp4
MOT17-14.mp4
MOT17-04.BOT-SORT-YOLOv7.COCO.mp4

Tracking performance

Results on MOT17 challenge test set

Tracker MOTA IDF1 HOTA
BoT-SORT 80.6 79.5 64.6
BoT-SORT-ReID 80.5 80.2 65.0

Results on MOT20 challenge test set

Tracker MOTA IDF1 HOTA
BoT-SORT 77.7 76.3 62.6
BoT-SORT-ReID 77.8 77.5 63.3

Installation

The code was tested on Ubuntu 20.04

BoT-SORT code is based on ByteTrack and FastReID.
Visit their installation guides for more setup options.

Setup with Anaconda

Step 1. Create Conda environment and install pytorch.

conda create -n botsort_env python=3.7
conda activate botsort_env

Step 2. Install torch and matched torchvision from pytorch.org.
The code was tested using torch 1.11.0+cu113 and torchvision==0.12.0

Step 3. Install BoT-SORT.

git clone https://github.com/NirAharon/BoT-SORT.git
cd BoT-SORT
pip3 install -r requirements.txt
python3 setup.py develop

Step 4. Install pycocotools.

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

Step 5. Others

# Cython-bbox
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 in the following structure:

<dataets_dir>
      β”‚
      β”œβ”€β”€ MOT17
      β”‚      β”œβ”€β”€ train
      β”‚      └── test    
      β”‚
      └── MOT20
             β”œβ”€β”€ train
             └── test

For training the ReID, detection patches must be generated as follows:

cd <BoT-SORT_dir>

# For MOT17 
python3 fast_reid/datasets/generate_mot_patches.py --data_path <dataets_dir> --mot 17

# For MOT20
 python3 fast_reid/datasets/generate_mot_patches.py --data_path <dataets_dir> --mot 20

Link dataset to FastReID export FASTREID_DATASETS=<BoT-SORT_dir>/fast_reid/datasets. If left unset, the default is fast_reid/datasets

Model Zoo

Download and store the trained models in 'pretrained' folder as follow:

<BoT-SORT_dir>/pretrained
  • We used the publicly available ByteTrack model zoo trained on MOT17, MOT20 and ablation study for YOLOX object detection.

  • Ours trained ReID models can be downloaded from MOT17-SBS-S50, MOT20-SBS-S50.

  • For multi-class MOT use YOLOX or YOLOv7 trained on COCO (or any custom weights).

Training

Train the ReID Module

After generating MOT ReID dataset as described in the 'Data Preparation' section.

cd <BoT-SORT_dir>

# For training MOT17 
python3 fast_reid/tools/train_net.py --config-file ./fast_reid/configs/MOT17/sbs_S50.yml MODEL.DEVICE "cuda:0"

# For training MOT20
python3 fast_reid/tools/train_net.py --config-file ./fast_reid/configs/MOT20/sbs_S50.yml MODEL.DEVICE "cuda:0"

Refer to FastReID repository for addition explanations and options.

Tracking

By submitting the txt files produced in this part to MOTChallenge website and you can get the same results as in the paper.
Tuning the tracking parameters carefully could lead to higher performance. In the paper we apply ByteTrack's calibration.

  • Test on MOT17
cd <BoT-SORT_dir>
python3 tools/track.py <dataets_dir/MOT17> --default-parameters --with-reid --benchmark "MOT17" --eval "test" --fp16 --fuse
python3 tools/interpolation.py --txt_path <path_to_track_result>
  • Test on MOT20
cd <BoT-SORT_dir>
python3 tools/track.py <dataets_dir/MOT20> --default-parameters --with-reid --benchmark "MOT20" --eval "test" --fp16 --fuse
python3 tools/interpolation.py --txt_path <path_to_track_result>
  • Evaluation on MOT17 validation set (the second half of the train set)
cd <BoT-SORT_dir>

# BoT-SORT
python3 tools/track.py <dataets_dir/MOT17> --default-parameters --benchmark "MOT17" --eval "val" --fp16 --fuse

# BoT-SORT-ReID
python3 tools/track.py <dataets_dir/MOT17> --default-parameters --with-reid --benchmark "MOT17" --eval "val" --fp16 --fuse
  • Other experiments

Other parameters can be used without passing --default-parameters flag.
For evaluating the train and validation sets we recommend using the official MOTChallenge evaluation code from TrackEval.

# For all the available tracking parameters, see:
python3 tools/track.py -h 
  • Experiments with YOLOv7

Other parameters can be used without passing --default-parameters flag.
For evaluating the train and validation sets we recommend using the official MOTChallenge evaluation code from TrackEval.

# For all the available tracking parameters, see:
python3 tools/track_yolov7.py -h 

Demo

Demo with BoT-SORT(-ReID) based YOLOX and multi-class.

cd <BoT-SORT_dir>

# Original example
python3 tools/demo.py video --path <path_to_video> -f yolox/exps/example/mot/yolox_x_mix_det.py -c pretrained/bytetrack_x_mot17.pth.tar --with-reid --fuse-score --fp16 --fuse --save_result

# Multi-class example
python3 tools/mc_demo.py video --path <path_to_video> -f yolox/exps/example/mot/yolox_x_mix_det.py -c pretrained/bytetrack_x_mot17.pth.tar --with-reid --fuse-score --fp16 --fuse --save_result

Demo with BoT-SORT(-ReID) based YOLOv7 and multi-class.

cd <BoT-SORT_dir>
python3 tools/mc_demo_yolov7.py --weights pretrained/yolov7-d6.pt --source <path_to_video/images> --fuse-score --agnostic-nms (--with-reid)

Note

Our camera motion compensation module is based on the OpenCV contrib C++ version of VideoStab Global Motion Estimation, which currently does not have a Python version.
Motion files can be generated using the C++ project called 'VideoCameraCorrection' in the GMC folder.
The generated files can be used from the tracker.

In addition, python-based motion estimation techniques are available and can be chosen by passing
'--cmc-method' <files | orb | ecc> to demo.py or track.py.

Citation

@article{aharon2022bot,
  title={BoT-SORT: Robust Associations Multi-Pedestrian Tracking},
  author={Aharon, Nir and Orfaig, Roy and Bobrovsky, Ben-Zion},
  journal={arXiv preprint arXiv:2206.14651},
  year={2022}
}

Acknowledgement

A large part of the codes, ideas and results are borrowed from ByteTrack, StorngSORT, FastReID, YOLOX and YOLOv7. Thanks for their excellent work!

About

BoT-SORT: Robust Associations Multi-Pedestrian Tracking

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 70.1%
  • Python 27.4%
  • C++ 2.1%
  • Cython 0.2%
  • CMake 0.1%
  • Dockerfile 0.1%