yieldloop / MMNet

PyTorch implementation of the paper ‘MMNet: Multi-Stage and Multi-Scale Fusion Network for RGB-D Salient Object Detection’ (ACMMM'2020)

MMNet-PyTorch (ACM MM, 2020)

MMNet: Multi-Stage and Multi-Scale Fusion Network for RGB-D Salient Object Detection. PDF

Update

Add performance of our MMNet trained on the new COME-Train dataset (8025 images):

https://pan.baidu.com/s/1S2ZT1AGqW0CfwaGFmubbbQ [code: wl4s]

Requirements

• pytorch 1.3.0+
• torchvision
• PIL
• Numpy

Testing

• Download the trained model from here [code: ofcn]
• Download test datasets from here [code: sva4]
• Modify your test_dataroot and test_datasets in test.py
• Test the MMNet: python test.py

Training

• Download the train-augment dataset from here [code: haxl]
• Download the pretrained backbone Res2Net(baseWidth = 48, scale = 2) from here
• Modify your train_dataroot and pre_trained_root in train.py
• Train the MMNet: python train.py

Results

• Saliency maps mentioned in the paper can be download from here [code: wl4s]
[1] The test_Results are obtained by trained on NJUD & NLPR & DUT (1485+700+800).
[2] The test_results_COME_train are obtained by trained on the new COME-Train dataset (8025).
• The saliency results can be evaluated by using the tool in Matlab

Citation

Please cite our paper if you use this repository in your reseach.

@inproceedings{MMNet20,   
author = {Liao, Guibiao and Gao, Wei and Jiang, Qiuping and Wang, Ronggang and Li, Ge},  
title = {MMNet: Multi-Stage and Multi-Scale Fusion Network for RGB-D Salient Object Detection},  
booktitle = {Proceedings of the 28th ACM International Conference on Multimedia},   
pages = {2436–2444},   
year = {2020}
}