The 1st place solution of track2 (Vehicle Re-Identification) in the NVIDIA AI City Challenge at CVPR 2021 Workshop.
Detailed information of NVIDIA AI City Challenge 2021 can be found here.
The code is modified from AICITY2020_DMT_VehicleReID, TransReID and reid_strong baseline.
-
cdto folder where you want to download this repo -
Run
git clone https://github.com/michuanhaohao/AICITY2021_Track2_DMT.git -
Install dependencies:
pip install requirements.txtWe use cuda 11.0/python 3.7/torch 1.6.0/torchvision 0.7.0 for training and testing.
-
Prepare Datasets Download Original dataset, Cropped_dataset, and SPGAN_dataset.
├── AIC21/
│ ├── AIC21_Track2_ReID/
│ ├── image_train/
│ ├── image_test/
│ ├── image_query/
│ ├── train_label.xml
│ ├── ...
│ ├── training_part_seg/
│ ├── cropped_patch/
│ ├── cropped_aic_test
│ ├── image_test/
│ ├── image_query/
│ ├── AIC21_Track2_ReID_Simulation/
│ ├── sys_image_train/
│ ├── sys_image_train_tr/- Put pre-trained models into ./pretrained/
We utilize 1 GPU (32GB) for training. You can train and test one backbone as follow.
# ResNext101-IBN-a
python train.py --config_file configs/stage1/resnext101a_384.yml MODEL.DEVICE_ID "('0')"
python train_stage2_v1.py --config_file configs/stage2/resnext101a_384.yml MODEL.DEVICE_ID "('0')" OUTPUT_DIR './logs/stage2/resnext101a_384/v1'
python train_stage2_v2.py --config_file configs/stage2/resnext101a_384.yml MODEL.DEVICE_ID "('0')" OUTPUT_DIR './logs/stage2/resnext101a_384/v2'
python test.py --config_file configs/stage2/1resnext101a_384.yml MODEL.DEVICE_ID "('0')" TEST.WEIGHT './logs/stage2/resnext101a_384/v1/resnext101_ibn_a_2.pth' OUTPUT_DIR './logs/stage2/resnext101a_384/v1'
python test.py --config_file configs/stage2/resnext101a_384.yml MODEL.DEVICE_ID "('0')" TEST.WEIGHT './logs/stage2/resnext101a_384/v2/resnext101_ibn_a_2.pth' OUTPUT_DIR './logs/stage2/resnext101a_384/v2'You should train camera and viewpoint models before the inference stage. You also can directly use our trained results (track_cam_rk.npy and track_view_rk.npy):
python train_cam.py --config_file configs/camera_view/camera_101a.yml
python train_view.py --config_file configs/camera_view/view_101a.ymlYou can train all eight backbones by checking run.sh. Then, you can ensemble all results:
python ensemble.pyAll trained models can be downloaded from here
| TeamName | mAP | Link |
|---|---|---|
| DMT(Ours) | 0.7445 | code |
| NewGeneration | 0.7151 | code |
| CyberHu | 0.6550 | code |
If you find our work useful in your research, please consider citing:
@inproceedings{luo2021empirical,
title={An Empirical Study of Vehicle Re-Identification on the AI City Challenge},
author={Luo, Hao and Chen, Weihua and Xu Xianzhe and Gu Jianyang and Zhang, Yuqi and Chong Liu and Jiang Qiyi and He, Shuting and Wang, Fan and Li, Hao},
booktitle={Proc. CVPR Workshops},
year={2021}
}