MSBA: Multiple Scales, Branches and Attention Network with Bag of Tricks for Person Re-Identification

Hanlin Tan ; Huaxin Xiao ; Xiaoyu Zhang ; Bin Dai ; Shiming Lai ; Yu Liu ; Maojun Zhang

National University of Defense Technology, China

This repository contains code for our paper MSBA: Multiple Scales, Branches and Attention Network with Bag of Tricks for Person Re-Identification, in which we have achieved the state-of-the-art performance on Market1501, DukeMTMC-reid, CUHK03(Detected) and MSMT17 datasets as of Feb, 2020. Please note the "state-of-the-art" is under restriction of no inference tricks and no additional information other than image contents. Namely, the results in the paper is without re-ranking or flipping, and the model does not take advantage of temporal information. Besides, the computation complexity of the network backbone should not be greater than Resnet-50.

The model proposed by the paper is named MSBA (Multiple Scales, Branches and Attention) network. Details and tricks are available at our paper.

Performance

• Performance of this repo:

Datasets	Model	Image Size	mAP	R1	Score	mAP (Re-rank+flip)	R1 (Re-rank+flip)	Score (Re-rank+flip)
market1501	MSBA-a	384x128	89.0	95.8	92.4	92.5	95.7	94.7
DukeMTMC-reID	MSBA-a	384x128	79.8 80.2	89.7 90.8	84.8 85.5	86.2	91.1	88.7
CUHK03(Detected)	MSBA-b	384x128	75.9	78.5	77.2	88.9	87.7	88.3
MSMT17	MSBA-a	384x128	59.0	75.3	67.1	66.2	77.5	71.8
MSMT17	MSBA-a	384x192	60.2	76.1	68.2	67.2	77.8	72.5

where score=(mAP+R1)/2.

• Performance comparison with other works mainly in 2019.

Training

• Clone the repo:

git clone --depth 1 https://github.com/TomHeaven/msba_reid
cd msba_reid

• Compile Cython codes:

cd csrc/eval_cylib; make

• Prepare the datasets at ../data folder and set the path by updating _C.DATASETS.ROOT_DIR of config/defaults.py .
• Prepare pretrained IBN-resnet50 weights from this repo and place it at ~/.torch/checkpoints/resnet50-ibn-a.pth.
• Run

sh scripts/train_resnet50_market1501.sh

to train and evaluate on market1501 dataset at resolution 384x128. Scripts for other datasets are available at scripts. The trained weights and logs are available at logs by default.

Inference using Pretrained Weights

• Download pretrained weights from GoogleDrive or BaiduYun (password:gizs) and place them at logs folder.
• Inference on Market1501 dataset using MSBA-a under resolution 384x128

python3 tools/val.py -cfg='configs/softmax_triplet.yml' \
DATASETS.TEST_NAMES 'market1501' \
MODEL.NAME 'resnet50_ibn' \
MODEL.BACKBONE 'resnet50' \
MODEL.WITH_IBN 'True' \
TEST.WEIGHT 'logs/384x128_MSBA-a/cuhk03/resnet50_ibn_bs64/ckpts/model_best.pth' \
TEST.IMS_PER_BATCH '128' \
INPUT.SIZE_TRAIN '[384, 128]' \
INPUT.SIZE_TEST '[384, 128]' 

• Inference on DukeMTMC-reID dataset using MSBA-a under resolution 384x128

python3 tools/val.py -cfg='configs/softmax_triplet.yml' \
DATASETS.TEST_NAMES 'dukemtmc' \
MODEL.NAME 'resnet50_ibn' \
MODEL.BACKBONE 'resnet50' \
MODEL.WITH_IBN 'True' \
TEST.WEIGHT 'logs/384x128_MSBA-a/dukemtmc/resnet50_ibn_bs64/ckpts/model_best.pth' \
TEST.IMS_PER_BATCH '128' \
INPUT.SIZE_TRAIN '[384, 128]' \
INPUT.SIZE_TEST '[384, 128]' 

• Inference on CUHK03(Detected) dataset using MSBA-b under resolution 384x128

python3 tools/val.py -cfg='configs/softmax_triplet.yml' \
DATASETS.TEST_NAMES 'cuhk03' \
MODEL.NAME 'resnet50_ibn' \
MODEL.BACKBONE 'resnet50' \
MODEL.WITH_IBN 'True' \
TEST.WEIGHT 'logs/384x128_MSBA-b/cuhk03/resnet50_ibn_bs64/ckpts/model_best.pth' \
TEST.IMS_PER_BATCH '128' \
INPUT.SIZE_TRAIN '[384, 128]' \
INPUT.SIZE_TEST '[384, 128]' 

• Inference on MSMT17 dataset using MSBA-a under resolution 384x128

python3 tools/val.py -cfg='configs/softmax_triplet.yml' \
DATASETS.TEST_NAMES 'msmt17' \
MODEL.NAME 'resnet50_ibn' \
MODEL.BACKBONE 'resnet50' \
MODEL.WITH_IBN 'True' \
TEST.WEIGHT 'logs/384x128_MSBA-a/msmt17/resnet50_ibn_bs64/ckpts/model_best.pth' \
TEST.IMS_PER_BATCH '128' \
INPUT.SIZE_TRAIN '[384, 128]' \
INPUT.SIZE_TEST '[384, 128]' 

The parameter TEST.WEIGHT specifies path of the pretrained weight. The parameter INPUT.SIZE_TEST controls the image size for inference. The file tools/val.py will also report performance results with re-ranking and flipping inference, which generally improves performance metrics by a large margin.

Switch between MSBA-a and MSBA-b

MSBA-b achieves better performance on CUHK03(Detected) dataset. Switch to MSBA-b by

cp modeling/baseline_parts_b.py  modeling/baseline_parts.py 

and back to MSBA-a by

cp modeling/baseline_parts_a.py  modeling/baseline_parts.py 

The difference is only in construction of loss function and affects training only. That is to say, you don't have to do this if you do inference only.

Optimized Re-ranking Implementation

• Optimized re-ranking. utils/re-ranking.py contains a heavily optimized re-ranking implementation: The distance matrix computation is done on GPU and the memory usage is heavily optimized compared with the original version. This is done when we join in a person ReID competition.

Reference

If you find our paper or repo helpful to your work, please cite our paper

@ARTICLE{tan2020msba, 
author={H. {Tan} and H. {Xiao} and X. {Zhang} and B. {Dai} and S. {Lai} and Y. {Liu} and M. {Zhang}
journal={IEEE Access}, 
title={MSBA: Multiple Scales, Branches and Attention Network With Bag of Tricks for Person Re-Identification}, year={2020}, volume={8}, 
pages={63632-63642},
}