OAFormer: Occlusion Aware Transformer for Camouflaged Object Detection

Authors: Xin Yang, Hengliang Zhu, Guojun Mao, Shuli Xing

This official repository contains the source code, prediction results, and evaluation toolbox of paper 'OAFormer: Occlusion Aware Transformer for Camouflaged Object Detection'. The technical report could be found at arXiv.

Figure 1: The overall architecture of the proposed occlusion aware transformer network. In part (a), we present the overall architecture of OAFormer. Part (b) displays the internal structures of the proposed neighborhood searching module (NSM).

1. Code

1.1. Requirements

Python 3.7
Pytorch 1.13.0
OpenCV 4.6.0.66
Timm 0.6.11
Numpy 1.23.4
Tqdm 4.64.1
TensorboardX 2.5.1
Apex 0.1

1.2. Download Datasets and Checkpoints.

• Datasets:

By default, you can put datasets into the folder 'dataset'.

• Checkpoints:

By default, you can put datasets into the folder 'checkpoint'.

OAFormer: Google Drive Backbone: Google Drive

1.3. Train.

python train.py

1.4. Test.

python test.py

1.5. Eval.

python eval.py

2. Proposed OAFormer

2.1 COD Benchmark Results:

The prediction of our OAFormer can be found in Google Drive. Here are quantitative performance comparison.

Acknowlegement

Thanks Zhirui Zhao providing a friendly codebase for binary segmentation tasks. And our code is built based on it.

Reference

You may want to cite:

@inproceedings{yang2023oaformer,
  title={OAFormer: Occlusion Aware Transformer for Camouflaged Object Detection},
  author={Yang, Xin and Zhu, Hengliang and Mao, Guojun and Xing, Shuli},
  booktitle={2023 IEEE International Conference on Multimedia and Expo (ICME)},
  pages={1421--1426},
  year={2023},
  organization={IEEE}
}

License

Code in this repo is for non-commercial use only.