Source code for "MFNet: Multi-filter Directive Network for Weakly Supervised Salient Object Detection", accepted in ICCV-2021, poster.
Yongri Piao, Jian Wang, Miao Zhang and Huchuan Lu. IIAU-OIP Lab.
- Windows 10
- Torch 1.8.1
- CUDA 10.0
- Python 3.7.4
- other environment requirment can be found in requirments.txt
link: https://pan.baidu.com/s/1omTCChQFWwNFhQ79AVD8rg. code: oipw
link: https://pan.baidu.com/s/1PBzDP1Hnf3RIvpARmxn2yA. code: oipw
Case1: please refer to this repository.
Case2: We also upload ready-made pseudo labels in Training data (the link above), you can directly use our offered two kinds of pseudo labels for convenience. CAMs are also presented if you needed.
MF_code -- data -- DUTS-Train -- image -- 10553 samples
-- ECSSD (not necessary)
-- pseudo labels -- label0_0 -- 10553 pseudo labels
-- label1_0 -- 10553 pseudo labels
Run main.py
Here you can set ECCSD dataset as validation set for optimal results by setting --val
to True
, of course it is not necessary in our work.
Run test_code.py
You need to configure your desired testset in --test_root
. Here you can also perform PAMR and CRF on saliency maps for a furthur refinements if you want, by setting --pamr
and --crf
to True. Noting that the results in our paper do not adopt these post-process for a fair comparison.
The evaluation code can be found in here.
We offer our saliency maps and checkpoints on various backbones (including DenseNet-169, ResNet-101, ResNet-50 and VGG-16) for more convenient comparison in the future. The results in our paper are all come from the model based on DenseNet-169, and we also highly recommend the following researchers adopt same backbone for a more fair and convenient comparison.
link: https://pan.baidu.com/s/1IRTEaEicYaCJ2TYjZV1lZA. code: oipw
link: https://pan.baidu.com/s/1lQm-MY4uwZZTW5NJTTXeyA. code: oipw
If you have any questions, please contact me: [[email protected]].
Thanks to pioneering helpful works: