This is a repo for independent learning. Tons of code was borrowed from ELECTRICITY-MTMC and VOC-ReID

What I did was to slightly improve its performance(measured by IDF1) by about 5%.

cd /ELECTRICITY-MTMC/
bash test_env.sh
bash ./train.sh;
bash ./test_ruiyan.sh

Below is the README of ELECTRICITY-MTMC.

@InProceedings{Qian_2020_CVPR_Workshops,
author = {Qian, Yijun and Yu, Lijun and Liu, Wenhe and Hauptmann, Alexander G.},
title = {ELECTRICITY: An Efficient Multi-Camera Vehicle Tracking System for Intelligent City},
booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2020}
}

Paper Linkage

Overview

We release the code for our winnning model on AI City 2020 Challenge (https://www.aicitychallenge.org/) Track 3. For more information please refer to our accepted paper in CVPR 2020 AI City Workshop.

Project Download

Firstly please download the project through:

git clone https://github.com/KevinQian97/ELECTRICITY-MTMC.git

Prerequisites

The code is built with many libraries, we have listed the official sites of part of them. If you encounter problems about the dependencies, please resort to these official sites for help.

• PyTorch 1.5 (we did not try other versions of pytorch)
• TensorboardX
• Opencv
• Cython
• AV
• ANACONDA

We have prepared the environment config file and suggest build the environment through ANACONDA.

cd ELECTRICITY-MTMC
conda env create -f environment.yml 
conda activate aic20_track3

Data Preparation

If you want to reproduce our results on AI City Challenge or train the model by yourself, please download the data set from: (https://www.aicitychallenge.org/2020-data-and-evaluation/) and put it under the folder datasets. Make sure the data structure is like:

• ELECTRICITY-MTMC
  • datasets
    • aic_20_trac3
      • test (test folder)
      • eval
      • validation (validation folder)
      • cam_timestamp
      • cam_loc
      • cam_framenum
      • train (train folder)

Pretrained Models

We also provided the pretrained model: Notice:The accuracy and map here is calculated on our inner split of validation set.

model	Acc 1	MAP	Epochs	Linkage
Agg_ResNet101	92.0%	82.3%	10	link

Inference

If you just want inference or reproduce our results, you can directly download our pretrained model and:

cd ELECTRICITY-MTMC
mkdir models
cd models
mkdir resnet101-Aic

Then put the pretrained model under this folder and run:

cd ELECTRICITY-MTMC
bash test.sh

The final results will locate at path ./exp/track3.txt

Training

If you want to train the model by yourself, please first generate training sets through:

bash ./prepare.sh

Then run:

bash ./train.sh

You will get trained model under path ./models/resnet101-Aic Finally run:

bash ./test.sh

The final results will locate at path ./exp/track3.txt

Performance

The speed is tested on four 2080Ti GPUs.

• Speed: 0.345x real-time (it means we only need 0.345 second to tackle 1 second video)
• IDF1: 0.4616

License

See LICENSE. Please read before use.