TransReID: Transformer-based Object Re-Identification

TransReID

The official repository for TransReID: Transformer-based Object Re-Identification achieves state-of-the-art performances on object re-ID, including person re-ID and vehicle re-ID.

Pipeline

Abaltion Study of Transformer-based Strong Baseline

Requirements

Installation

pip install -r requirements.txt
(we use /torch 1.6.0 /torchvision 0.7.0 /timm 0.3.2 /cuda 10.1 / 16G or 32G V100 for training and evaluation.
Note that we use torch.cuda.amp to accelerate speed of training which requires pytorch >=1.6)

Prepare Datasets

mkdir data

Download the person datasets Market-1501, MSMT17, DukeMTMC-reID,Occluded-Duke, and the vehicle datasets VehicleID, VeRi-776,
Then unzip them and rename them under the directory like

data
├── market1501
│   └── images ..
├── MSMT17
│   └── images ..
├── dukemtmcreid
│   └── images ..
├── Occluded_Duke
│   └── images ..
├── VehicleID_V1.0
│   └── images ..
└── VeRi
    └── images ..

Prepare DeiT or ViT Pre-trained Models

You need to download the ImageNet pretrained transformer model : ViT-Base, ViT-Small, DeiT-Small, DeiT-Base

Training

We utilize 1 GPU for training.

python train.py --config_file configs/transformer_base.yml MODEL.DEVICE_ID "('your device id')" MODEL.STRIDE_SIZE ${1} MODEL.SIE_CAMERA ${2} MODEL.SIE_VIEW ${3} MODEL.JPM ${4} MODEL.TRANSFORMER_TYPE ${5} OUTPUT_DIR ${OUTPUT_DIR} DATASETS.NAMES "('your dataset name')"

Arguments

• ${1}: stride size for pure transformer, e.g. [16, 16], [14, 14], [12, 12]
• ${2}: whether using SIE with camera, True or False.
• ${3}: whether using SIE with view, True or False.
• ${4}: whether using JPM, True or False.
• ${5}: choose transformer type from 'vit_base_patch16_224_TransReID',(The structure of the deit is the same as that of the vit, and only need to change the imagenet pretrained model) 'vit_small_patch16_224_TransReID','deit_small_patch16_224_TransReID',
• ${OUTPUT_DIR}: folder for saving logs and checkpoints, e.g. ../logs/market1501

or you can directly train with following yml and commands:

# DukeMTMC transformer-based baseline
python train.py --config_file configs/DukeMTMC/vit_base.yml MODEL.DEVICE_ID "('0')"
# DukeMTMC baseline + JPM
python train.py --config_file configs/DukeMTMC/vit_jpm.yml MODEL.DEVICE_ID "('0')"
# DukeMTMC baseline + SIE
python train.py --config_file configs/DukeMTMC/vit_sie.yml MODEL.DEVICE_ID "('0')"
# DukeMTMC TransReID (baseline + SIE + JPM)
python train.py --config_file configs/DukeMTMC/vit_transreid.yml MODEL.DEVICE_ID "('0')"
# DukeMTMC TransReID with stride size [12, 12]
python train.py --config_file configs/DukeMTMC/vit_transreid_stride.yml MODEL.DEVICE_ID "('0')"

# MSMT17
python train.py --config_file configs/MSMT17/vit_transreid_stride.yml MODEL.DEVICE_ID "('0')"
# OCC_Duke
python train.py --config_file configs/OCC_Duke/vit_transreid_stride.yml MODEL.DEVICE_ID "('0')"
# Market
python train.py --config_file configs/Market/vit_transreid_stride.yml MODEL.DEVICE_ID "('0')"
# VeRi
python train.py --config_file configs/VeRi/vit_transreid_stride.yml MODEL.DEVICE_ID "('0')"

# VehicleID (The dataset is large and we utilize 4 v100 GPUs for training )
CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 --master_port 66666 train.py --config_file configs/VehicleID/vit_transreid_stride.yml MODEL.DIST_TRAIN True
#  or using following commands:
Bash dist_train.sh 

Tips: For person datasets with size 256x128, TransReID with stride occupies 12GB GPU memory and TransReID occupies 7GB GPU memory.

Evaluation

python test.py --config_file 'choose which config to test' MODEL.DEVICE_ID "('your device id')" TEST.WEIGHT "('your path of trained checkpoints')"

Some examples:

# DukeMTMC
python test.py --config_file configs/DukeMTMC/vit_transreid_stride.yml MODEL.DEVICE_ID "('0')"  TEST.WEIGHT '../logs/duke_vit_transreid_stride/transformer_120.pth'
# MSMT17
python test.py --config_file configs/MSMT17/vit_transreid_stride.yml MODEL.DEVICE_ID "('0')" TEST.WEIGHT '../logs/msmt17_vit_transreid_stride/transformer_120.pth'
# OCC_Duke
python test.py --config_file configs/OCC_Duke/vit_transreid_stride.yml MODEL.DEVICE_ID "('0')" TEST.WEIGHT '../logs/occ_duke_vit_transreid_stride/transformer_120.pth'
# Market
python test.py --config_file configs/Market/vit_transreid_stride.yml MODEL.DEVICE_ID "('0')"  TEST.WEIGHT '../logs/market_vit_transreid_stride/transformer_120.pth'
# VeRi
python test.py --config_file configs/VeRi/vit_transreid_stride.yml MODEL.DEVICE_ID "('0')" TEST.WEIGHT '../logs/veri_vit_transreid_stride/transformer_120.pth'

# VehicleID (We test 10 times and get the final average score to avoid randomness)
python test.py --config_file configs/VehicleID/vit_transreid_stride.yml MODEL.DEVICE_ID "('0')" TEST.WEIGHT '../logs/vehicleID_vit_transreid_stride/transformer_120.pth'

Trained Models and logs (Size 256)

Note: We reorganize code and the performances are slightly different from the paper's.

Acknowledgement

Codebase from reid-strong-baseline , pytorch-image-models

We import veri776 viewpoint label from repo: https://github.com/Zhongdao/VehicleReIDKeyPointData

Citation

If you find this code useful for your research, please cite our paper

@article{he2021transreid,
  title={TransReID: Transformer-based Object Re-Identification},
  author={He, Shuting and Luo, Hao and Wang, Pichao and Wang, Fan and Li, Hao and Jiang, Wei},
  journal={arXiv preprint arXiv:2102.04378},
  year={2021}
}

Contact

If you have any question, please feel free to contact us. E-mail: [email protected] , [email protected]

GitHub

https://github.com/heshuting555/TransReID