Visual Attention Network (VAN) for Detection

This repo is a PyTorch implementation applying VAN (Visual Attention Network) to 2D Object Detection. Our implementation is mainly based on VAN-Segmentation and MMDetection. Thanks to the authors.

More details about the VAN can be found in Visual Attention Network.

Citation

@article{guo2022visual,
  title={Visual Attention Network},
  author={Guo, Meng-Hao and Lu, Cheng-Ze and Liu, Zheng-Ning and Cheng, Ming-Ming and Hu, Shi-Min},
  journal={arXiv preprint arXiv:2202.09741},
  year={2022}
}

Preparation

Install MMDetection and download COCO according to the guidelines in MMDetection.

Requirements

We recommend following the official instructions for installing the Open-MMLab libraries, using mim. Otherwise, version mismatches are likely.

pip install wandb timm pycocotools openmim

mim install mmcv-full==1.7.1 mmdet==2.27.0

As getting the correct sets of versions correct can be tricky, we provide the exact enviroments used in our tests in the conda-freeze.txt respective pip-freeze.txt. We used our own fork of MMDetection for our adapted copy-paste mechanism and our evaluation. (@commit 9e62e9b4f05aedf5b0b28e5c7619ef8e89097cc1)

Training

We use 3 GPUs for training by default. Run:

./dist_train.sh /path/to/config 3

Evaluation

To evaluate the model, run:

./dist_test.sh /path/to/config /path/to/checkpoint_file 3 --eval bbox

FLOPs

Install torchprofile using

pip install torchprofile

To calculate FLOPs for a model, run:

bash tools/flops.sh /path/to/config --shape 1333 800

In our evaluation, we used the updated analysis tools of our MMDetection fork.

Acknowledgment

Our implementation is mainly based on VAN-Segmentation and MMDetection. Thanks to the authors.

LICENSE

This repo is under the Apache-2.0 license. For commercial use, please contact the authors.