Skip to content

heshuting555/AICITY2020_DMT_VehicleReID

master
Switch branches/tags
Code

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Multi-Domain Learning and Identity Mining for Vehicle Re-Identification

This repository contains our source code of Track2 in the NVIDIA AI City Challenge at CVPR 2020 Workshop. Our paper

Authors

Introduction

Detailed information of NVIDIA AI City Challenge 2020 can be found here.

The code is modified from reid_strong baseline and person_reid_tiny_baseline.

Get Started

  1. cd to folder where you want to download this repo

  2. Run git clone https://github.com/heshuting555/AICITY2020_DMT_VehicleReID.git

  3. Install dependencies:

    We use cuda 9.0/python 3.6.7/torch 1.2.0/torchvision 0.4.0 for training and testing.

  4. Prepare dataset. we have to change the first line in AIC20_track2/AIC20_ReID/train_label.xml as below:

    <?xml version="1.0" encoding="gb2312"?>

    into

    <?xml version="1.0" encoding="utf-8"?>
  5. ResNet-ibn is applied as the backbone. Download ImageNet pretrained model here

RUN

  1. If you want to get the same score as online in the AI City Challenge 2020 Track2. Use the following commands:

    bash run.sh

    Note: you can download our trained model and Distance matrix in the AICITY2020 here

  2. If you want to use our Multi-Domain Learning.

    # you need to train a model in a Multi-Domain Datasets first.(E.g: you can add simulation datasets to aic and then test on the aic)
    
    python train.py --config_file='configs/baseline_aic_finetune.yml' MODEL.PRETRAIN_PATH "('your path for trained checkpoints')" MODEL.DEVICE_ID "('your device id')" OUTPUT_DIR "('your path to save checkpoints and logs')"
  3. If you want to try our Identity Mining.

    # First, genereate the selected query ids
    
    python test_mining.py --config_file='configs/test_identity_mining.yml'  TEST.WEIGHT "('your path for trained checkpoints')" OUTPUT_DIR "('your path to save selected query id')" --thresh 0.49

    Note: The quality of the query id depends on the performance of TEST.WEIGHT. And you can change the value of thresh to get more or less query ids.

    # Then,  train the model with trainset and testset(selected by the above selected query id)
    
    python train_IM.py --config_file='configs/baseline_aic.yml'
    --config_file_test='configs/test_train_IM.yml' OUTPUT_DIR "('your path to save checkpoints and logs')" MODEL.THRESH "(0.23)"

    Note: you can change the value of MODEL.THRESH which determines how many test sets added to the train sets.

  4. If you want to generate crop images please refer to crop_dataset_generate directory for detail.

  5. You can visualize the result given a track2.txt result (AICITY required submission format).

    python vis_txt_result.py --base_dir ('your path to the datasets') --result ('result file (txt format) path')
  6. If you want to use our baseline on public datasets (such as VeRi datasets).

    python train.py --config_file='configs/baseline_veri_r50.yml' MODEL.DEVICE_ID "('your device id')" OUTPUT_DIR "('your path to save checkpoints and logs')"

Results (mAP/Rank1)

Model AICITY2020
Resnet101_ibn_a (baseline) 59.73/69.30
+ Multi-Domain Learning 65.25/71.96
+ Identity Mining 68.54/74.81
+ Ensemble 73.22/80.42
Backbone (baseline) VeRi download
ResNet50 (batch 48) 79.8/95.0 model | log
Resnet50_ibn_a (batch 48) 81.4/96.5 model | log
Resnet101_ibn_a (batch 48) 82.8/97.1 model | log

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{he2020multi,
 title={Multi-Domain Learning and Identity Mining for Vehicle Re-Identification},
 author={He, Shuting and Luo, Hao and Chen, Weihua and Zhang, Miao and Zhang, Yuqi and Wang, Fan and Li, Hao and Jiang, Wei},
 booktitle={Proc. CVPR Workshops},
 year={2020}
}

About

The 3rd Place Submission to AICity Challenge 2020 Track2

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published