M³Net: Multilevel, Mixed and Multistage Attention Network for Salient Object Detection

Source code of 'M³Net: Multilevel, Mixed and Multistage Attention Network for Salient Object Detection'. paper link

Environment

Python 3.9.13 and Pytorch 1.11.0. Details can be found in requirements.txt.

Data Preparation

All datasets used can be downloaded at here [arrr].

Training set

We use the training set of DUTS to train our M³Net.

Testing Set

We use the testing set of DUTS, ECSSD, HKU-IS, PASCAL-S, DUT-O, and SOD to test our M³Net. After Downloading, put them into /datasets folder.

Your /datasets folder should look like this:

-- datasets
   |-- DUT-O
   |   |--imgs
   |   |--gt
   |-- DUTS-TR
   |   |--imgs
   |   |--gt
   |-- ECSSD
   |   |--imgs
   |   |--gt
   ...

Training and Testing

1. Download the pretrained backbone weights and put it into pretrained_model/ folder. ResNet [uxcz], SwinTransformer, T2T-ViT, EfficientNet are currently supported.

2. Run python train_test.py --train True --test True --record='record.txt' for training and testing. The predictions will be in preds/ folder and the training records will be in record.txt file.

Evaluation

Pre-calculated saliency maps: M³Net-R [uqsr], M³Net-S [6jyh]
Pre-trained weights: M³Net-R [m789], M³Net-S [4wnw]

For PR curve and F curve, we use the code provided by this repo: [BASNet, CVPR-2019].
For MAE, Weighted F measure, E score and S score, we use the code provided by this repo: [PySODMetrics].

For more information about evaluation, please refer to Evaluation/Guidance.md.

Evaluation Results

Quantitative Evaluation

Precision-recall and F-measure curves

Visual Comparison

Acknowledgement

Our idea is inspired by VST and MiNet. Thanks for their excellent works. We also appreciate the data loading and enhancement code provided by plemeri, as well as the efficient evaluation tool provided by lartpang.

Citation

If you think our work is helpful, please cite

@misc{yuan2023m3net,
      title={M$^3$Net: Multilevel, Mixed and Multistage Attention Network for Salient Object Detection}, 
      author={Yao Yuan and Pan Gao and XiaoYang Tan},
      year={2023},
      eprint={2309.08365},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}