Attention-in-Attention-Networks

Implementation of ICCV19: Bilinear attention networks for person retrieval and its extension for Journal, which is to be appeared in TPAMI: Attention in attention networks for person retrieval.

If you find this code useful in your research then please cite

@InProceedings{Fang_2019_ICCV,
author = {Fang, Pengfei and Zhou, Jieming and Roy, Soumava Kumar and Petersson, Lars and Harandi, Mehrtash},
title = {Bilinear Attention Networks for Person Retrieval},
booktitle = {ICCV},
month = {October},
year = {2019}
}

@InProceedings{Fang_2021_PAMI,
author = {Fang, Pengfei and Zhou, Jieming and Roy, Soumava Kumar and and Pan Ji and Petersson, Lars and Harandi, Mehrtash},
title = {Attention in Attention Networks for Person Retrieval},
booktitle = {TPAMI},
year = {2021}
}

Setups

Python 2.7.16
torch 1.1.0
torchvision 0.3.0

Datasets

CUHK01
CUHK03
Market-1501
DukeMTMC
MSMT17

Networks

Inception
Inception + part feature extractor
other networks...

Training network on benchmark datasets

Here is an example-step one

python train_imgreid_xent_semitri.py --pretrained --resume 'pretrained_model/inceptionV1_bn_noModule.pth.tar' --height 256 --width 128 --max-epoch 50 --train-batch 64 --focused-parts 4 --factor-of-scale-factors 1 --stepsize 150 200 250 --gamma 0.1 --lr 0.0005 --weight-decay 0.0001 --drop-rate 0.1 --margin 1 --num_trip 10 --optim 'adam' --save-dir './market/log' --arch bilinear_baseline --gpu-devices 0,1

step two

python train_imgreid_xent_semitri_re.py --resume './market/log/checkpoint_ep50.pth.tar' --height 256 --width 128 --max-epoch 300 --train-batch 64 --focused-parts 4 --factor-of-scale-factors 1 --stepsize 150 200 250 --gamma 0.1 --lr 0.0005 --weight-decay 0.0001 --drop-rate 0.1 --margin 1 --num_trip 10 --optim 'adam' --save-dir './market/log_re' --arch bilinear_baseline --gpu-devices 0,1

Acknowledgement

Thank you for the codebases: deep-person-reid

mygit007hub / Attention-in-Attention-Networks