SCSA

This repo is the official of implementation of "SCSA: Exploring the Synergistic Effects Between Spatial and Channel Attention".

Introduction

In this paper, starting from the synergy of multi-semantic information, we propose a plug-and-play Spatial and Channel Synergistic Attention module(SCSA).

We conduct extensive experiments on seven benchmark datasets, including classification on ImageNet-1K, object detection on MSCOCO 2017, segmentation on ADE20K, and four other complex scene detection datasets to validate the effectiveness of our method.

Running

Install

We implement SCSA using MMPretrain V1.2.0, MMDetection V3.3.0, MMSegmentation V1.2.2 and MMCV V2.1.0.
We train and test our models under python=3.10, pytorch=2.1.1, cuda=11.8.

# Create a virtual environment and activate it.
pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu118
pip install -U openmim
mim install -e .

Data preparation

The ImageNet dataset should be prepared as follows:

$ tree data
imagenet
├── train
│   ├── class1
│   │   ├── img1.jpeg
│   │   ├── img2.jpeg
│   │   └── ...
│   ├── class2
│   │   ├── img3.jpeg
│   │   └── ...
│   └── ...
└── val
    ├── class1
    │   ├── img4.jpeg
    │   ├── img5.jpeg
    │   └── ...
    ├── class2
    │   ├── img6.jpeg
    │   └── ...
    └── ...

Training

Our work employs a single GPU for training on classification tasks.
Here is an example: train SCSA + ResNet-50 with an single GPU:

python tools\train.py work_dirs\resnet50_1xb128_in1k_scsa.py --work-dir path_to_exp --amp 

Testing

Test SCSA + ResNet-50 with an single GPU:

python tools\test.py work_dirs\resnet50_1xb128_in1k_scsa.py path_to_checkpoint --work-dir path_to_exp

Results

We will open source the relevant model weights later.

Acknowledgement

The code in this repository is developed based on the MMPretrain. Furthermore, the detection and segmentation tasks involved in this work are implemented based on the MMDetection and MMSegmentation.

Cite SCSA

If you find this repository useful, please use the following BibTeX entry for citation.

@article{si2024SCSA,
  title={SCSA: Exploring the Synergistic Effects Between Spatial and Channel Attention},
  author={Si, Yunzhong and Xu, Huiying and Zhu, Xinzhong and Zhang, Wenhao and Dong, Yao and Chen, Yuxing and Li, Hongbo},
  journal={arXiv preprint arXiv:2407.05128},
  year={2024}
}

Concat

If you have any questions, please feel free to contact the authors.

Yunzhong Si: siyunzhong@zjnu.edu.cn