eraserNut / MTMT

Code for the CVPR 2020 paper "A Multi-task Mean Teacher for Semi-supervised Shadow Detection"

A Multi-task Mean Teacher for Semi-supervised Shadow Detection

by Zhihao Chen, Lei Zhu, Liang Wan, Song Wang, Wei Feng, and Pheng-Ann Heng [paper link]

News: In 2020.9.17, We release the unsorted code for other researchers. The sorted code will be released after.

---

Citation

@inproceedings{chen20MTMT,
     author = {Chen, Zhihao and Zhu, Lei and Wan, Liang and Wang, Song and Feng, Wei and Heng, Pheng-Ann},
     title = {A Multi-task Mean Teacher for Semi-supervised Shadow Detection},
     booktitle = {CVPR},
     year = {2020}
}

Shadow detection results at test datasets

The results of shadow detection(w & w/o crf) on three datasets (SBU, UCF, ISTD) can be found at Google Drive or BaiduNetdisk(password:131b for BaiduNetdisk).

Trained Model

You can download the trained model which is reported in our paper at BaiduNetdisk(password: h52i) or Google Drive.

Requirement

• Python 3.6
• PyTorch 1.3.1(After 0.4.0 would be ok)
• torchvision
• numpy
• tqdm
• PIL
• pydensecrf (here to install)

Training

1. Set ...
2. Set ...
3. Run by python train.py

The pretrained ResNeXt model is ported from the official torch version, using the convertor provided by clcarwin. You can directly download the pretrained model ported by me.

Testing

1. Set ...
2. Put ...
3. Run by python test_MT.py

Useful links

UCF dataset: Google Drive or BaiduNetdisk(password:o4ub for BaiduNetdisk)

SBU dataset: BaiduNetdisk(password:38qw for BaiduNetdisk)

Part of unlabel data that collected from internet: Google Drive or BaiduNetdisk(password: n1nb for BaiduNetdisk)