Awesome

CAViT

[ECCV 2022] CAViT: Contextual Alignment Vision Transformer for Video Object Re-identification

Preparation

• Download datasets

  • MARS
  • MARS_DL
  • LS_VID
  • PRID-2011
  • iLIDS-VID
  • vveri901

• add data path

  • run export FASTREID_DATASETS=/path/to/datasets/ in the terminal,
  • or add export FASTREID_DATASETS=/path/to/datasets/ to your ~/.bashrc

• Install requirement

conda create -n reid python=3.7
conda activate reid
conda install pytorch==1.6.0 torchvision tensorboard -c pytorch

# if you use A100 gpu, please install pytorch >= 1.7 and cudatoolkit >=11.0
# for example
# conda install pytorch==1.9.1 torchvision==0.10.1 torchaudio==0.9.1 cudatoolkit=11.3 -c pytorch -c conda-forge

pip install -r docs/requirements.txt

• using cython for evalution

cd fastreid/evaluation/rank_cylib
make all

Support backbone

• swin transformer & swin transformer 3D
• ViT & ViT 3D (timeformer)
• TSM
• AP3D
• BickNet
• token shift

Training

1. If you want to train with 4-GPU, run:

CUDA_VISIBLE_DEVICES=2 python3 projects/CAViT/train.py --config-file projects/CAViT/configs/cavit_prid2011.yml  --num-gpus 1

2. If you want to train with 4-GPU, run:

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 projects/CAViT/train.py --config-file projects/CAViT/configs/cavit_prid2011.yml  --num-gpus 4  HEADS.NORM syncBN

Testing

1. If you want to test with 1-GPU, run:

CUDA_VISIBLE_DEVICES=2 python3 projects/CAViT/train.py --config-file projects/CAViT/configs/cavit_prid2011.yml  --num-gpus 1  --eval-only MODEL.WEIGHTS logs/prid2011/model_best.pth TEMP.TEST.ALL True

2. If you want to test with 4-GPUs, run:

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 projects/CAViT/train.py --config-file projects/CAViT/configs/cavit_prid2011.yml  --num-gpus 4  --eval-only MODEL.WEIGHTS logs/prid2011/model_best.pth TEMP.TEST.ALL True  HEADS.NORM syncBN

This repo also implements the fixed length sequeence testing & the flexible length testing.

1. Flexible testing (use all frames in the sequence for testing)

CUDA_VISIBLE_DEVICES=2 python3 projects/CAViT/train.py --config-file projects/CAViT/configs/cavit_prid2011.yml  --num-gpus 1  --eval-only MODEL.WEIGHTS logs/prid2011/model_best.pth TEMP.TEST.ALL True

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 projects/CAViT/train.py --config-file projects/CAViT/configs/cavit_prid2011.yml  --num-gpus 4  --eval-only MODEL.WEIGHTS logs/prid2011/model_best.pth TEMP.TEST.ALL True  HEADS.NORM syncBN

2. Fixed length testing

CUDA_VISIBLE_DEVICES=2 python3 projects/CAViT/train.py --config-file projects/CAViT/configs/cavit_prid2011.yml  --num-gpus 1  --eval-only MODEL.WEIGHTS logs/prid2011/model_best.pth TEMP.TEST.ALL False  TEMP.TEST.SEQ_SIZE  8   TEMP.TEST.TRACK_SPLIT 128

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 projects/CAViT/train.py --config-file projects/CAViT/configs/cavit_prid2011.yml  --num-gpus 4  --eval-only MODEL.WEIGHTS logs/prid2011/model_best.pth HEADS.NORM syncBN TEMP.TEST.ALL False  TEMP.TEST.SEQ_SIZE  8   TEMP.TEST.TRACK_SPLIT 128

Notice

• The experimental results on these two data sets are unstable, particulary on PRID-2011. (e.g., Dropout in ViT, Diffent GPUs, pytorch versions, and etc., all of them may influence the final results.)

• If you train models with multi-gpus, please fix HEADS.NORM: BN to HEADS.NORM: syncBN in the config file. This may influence the results.

• For MARS_DL dataset, NewRank-1 and New_mAP are the performance metric, while on other datasets, Rank-1 ,Rank-5, and mAP are the performance metric.

• We reproduce the experiments of our paper with 1% performance gap. The model

Download models

• onrdrive
• baidu cloud (code 9218)

Contacts

If you have any question about the project, please feel free to contact me.

E-mail: jinlin.wu@nlpr.ia.ac.cn

ACKNOWLEDGEMENTS

• The code was developed based on the fast-reid toolbox https://github.com/JDAI-CV/fast-reid.
• Thanks for AP3D of Xinqian Gu providing video reid code base.