Implementation of Vehicle Re-Identification Based on Complementary Features for 2020 AICity Challenge Track2

This repository contains the source codes of vehicle Re-ID of our implementation for 2020 AICity Challenge, and we got 5-th place in the vehicle Re-ID track of AIC2020. Our paper

Dependencies

python 2.7 / python 3.6

pytorch 1.0 +

torchvision 0.2.1 +

metric_learn 0.5.0 +

cv2 3.0 + refer to our source codes for other dependencies

Datasets

Datasets used in our implementation are avialable at

Datasets	Description	Download link
Original train set	CityFlow train set and Simulation Set by VehicleX	link
Crop train set for cityflow	CityFlow train set cropped by 2019 1st Baidu's detector	link
All images with fake label test test	CityFlow train set & test set(fake label) and Simulation Set by VehicleX	link

Models

Notes: You can directly use our trained model and extracted features for implementation, the link is as follows:

Models	Usage	Description	Download link
ImageNet pretrained models	Train	for global_model/pretrain_models	link
ImageNet pretrained models	Train	for mgn_model/weights	link
Vehicle ReID models trained by us	Test	contain four models checkpoint mentioned in our paper	link

Features	Usage	Description	Download link
Pkls for features	Test	features extracted by each model and performed by several post-processing	link

Code Structure

Each part has its own README file.

• global_model contains source codes of training vehicle reid models including se_resnext101, se_resnet152, resnet152, hrnet_48w, se_resnet152_ibnb, densenet161, dpn107, senet154 etc.

• mgn_mode contains source codes of training vehicle reid models including resnet152 with MGN, resnet152 with SAC.

• post_processing contains several post-processing methods for Re-ID task.

Running Code orderly

1. Train each single model in global_model and mgn_model.

2. Extract features for test set using each single model in global_model and mgn_model.

3. Move all pkls of features to the same directory and utilize several post-processing methods to improve single model performence.

Detail steps to reproduce our result

Trainng

Train each single model in global_model and mgn_model.

see global_models/README.md and mgn_model/README.md for detailed training steps.

Testing

1 Download the vehicle ReID models trained by us from the google-drive(link).

2 unzip the downloaded file ckpt.tar in the current directory, and make a link in both global_model and mgn_model directory

3 download imagenet pretrained models(link) for global_model/pretrain_models and unzip in global_mddel directory

4 download imagenet pretrained models(link) for mgn_model/weights and unzip in mgn_mddel directory

5 enter global_model directory, modify query_path and gallery_path in extract_feature_val.sh

6 extract features using global models, the features will be saved in post_processing/val_pkl_final

sh extract_feature_val.sh

7 enter mgn_model directory, modify query_path and gallery_path in extract_feature_val.sh

8 extract features using mgn model, the features will be saved in post_processing/val_pkl_final

sh extract_feature_val.sh

9 enter post_processing directory , run

sh test_final.sh

you will get track2.txt which is the final result

---

Basic framework

• The basic framework of our approach .
  

• Results generated by our method.