Deep Sort with PyTorch

Update(1-1-2020)

Changes

fix bugs
refactor code
accerate detection by adding nms on gpu

Latest Update(07-22)

Changes

bug fix (Thanks @JieChen91 and @yingsen1 for bug reporting).
using batch for feature extracting for each frame, which lead to a small speed up.
code improvement.

Futher improvement direction

Train detector on specific dataset rather than the official one.
Retrain REID model on pedestrain dataset for better performance.
Replace YOLOv3 detector with advanced ones.

Any contributions to this repository is welcome!

Introduction

This is an implement of MOT tracking algorithm deep sort. Deep sort is basicly the same with sort but added a CNN model to extract features in image of human part bounded by a detector. This CNN model is indeed a RE-ID model and the detector used in PAPER is FasterRCNN , and the original source code is HERE.
However in original code, the CNN model is implemented with tensorflow, which I'm not familier with. SO I re-implemented the CNN feature extraction model with PyTorch, and changed the CNN model a little bit. Also, I use YOLOv3 to generate bboxes instead of FasterRCNN.

Dependencies

python 3 (python2 not sure)
numpy
scipy
opencv-python
sklearn
torch >= 0.4
torchvision >= 0.1
pillow
vizer
edict

Quick Start

Check all dependencies installed

pip install -r requirements.txt

for user in china, you can specify pypi source to accelerate install like:

pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple

Clone this repository

git clone git@github.com:ZQPei/deep_sort_pytorch.git

Download YOLOv3 parameters

cd detector/YOLOv3/weight/
wget https://pjreddie.com/media/files/yolov3.weights
wget https://pjreddie.com/media/files/yolov3-tiny.weights
cd ../../../

Download deepsort parameters ckpt.t7

cd deep_sort/deep/checkpoint
# download ckpt.t7 from
https://drive.google.com/drive/folders/1xhG0kRH1EX5B9_Iz8gQJb7UNnn_riXi6 to this folder
cd ../../../

Compile nms module

cd detector/YOLOv3/nms
sh build.sh
cd ../../..

Notice: If compiling failed, the simplist way is to **Upgrade your pytorch >= 1.1 and torchvision >= 0.3" and you can avoid the troublesome compiling problems which are most likely caused by either gcc version too low or libraries missing.

Run demo

usage: python yolov3_deepsort.py VIDEO_PATH
                                [--help]
                                [--frame_interval FRAME_INTERVAL]
                                [--config_detection CONFIG_DETECTION]
                                [--config_deepsort CONFIG_DEEPSORT]
                                [--ignore_display]
                                [--display_width DISPLAY_WIDTH]
                                [--display_height DISPLAY_HEIGHT]
                                [--save_path SAVE_PATH]          
                                [--cpu]          

# yolov3 + deepsort
python yolov3_deepsort.py [VIDEO_PATH]

# yolov3_tiny + deepsort
python yolov3_deepsort.py [VIDEO_PATH] --config_detection ./configs/yolov3_tiny.yaml

# yolov3 + deepsort on webcam
python3 yolov3_deepsort.py /dev/video0 --camera 0

# yolov3_tiny + deepsort on webcam
python3 yolov3_deepsort.py /dev/video0 --config_detection ./configs/yolov3_tiny.yaml --camera 0

If you dont support X server, use --ignore_display to disable display. Results will be saved to ./demo/demo.avi.

All files above can also be accessed from BaiduDisk!
linker：BaiduDisk passwd：fbuw

Training the RE-ID model

The original model used in paper is in original_model.py, and its parameter here original_ckpt.t7.

To train the model, first you need download Market1501 dataset or Mars dataset.

Then you can try train.py to train your own parameter and evaluate it using test.py and evaluate.py.

Demo videos and images

demo.avi demo2.avi

References

paper: Simple Online and Realtime Tracking with a Deep Association Metric
code: nwojke/deep_sort
paper: YOLOv3
code: Joseph Redmon/yolov3

banana16314/deep_sort_pytorch