Learning from Synthetic Shadows for Shadow Detection and Removal (IEEE TCSVT 2021)

Overview

This repo is for the paper "Learning from Synthetic Shadows for Shadow Detection and Removal". We present SynShadow, a novel large-scale synthetic shadow/shadow-free/matte image triplets dataset and pipeline to synthesize it. We further show how to use SynShadow for robust and efficient shadow detection and removal.

In this repo, we provide

• SynShadow dataset: ./datasets
• SP+M implementation: ./src
• Trained models and results: below

If you find this code or dataset useful for your research, please cite our paper:

@article{inoue2021learning,
  title={{Learning from Synthetic Shadows for Shadow Detection and Removal}},
  author={Inoue, Naoto and Yamasaki, Toshihiko},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  year={2021},
  volume={31},
  number={11},
  pages={4187-4197},
  doi={10.1109/TCSVT.2020.3047977}
}

Trained Models and Results

We provide the models for shadow detection and removal for convenience. Downloaded models should be placed under ./checkpoints.

Shadow Detection

ALl the results are in 480x640. BER is reported for 480x640 images. Below are results evaluated on ISTD test set. DSDNet++ is a modified variant of DSDNet.

Shadow Removal

ALl the results are in 480x640. For the pre-trained weights, we only provide SP+M weights, since this repository has full implementation of it. RMSE is reported for 480x640 images.

Model: SP+M

Model: DHAN

Note: we have accidentially removed some files and cannot provide some results.