Towards Real-World Shadow Removal With a Shadow Simulation Method and a Two-Stage Framework

Jianhao Gao, Quanlong Zheng, Yandong Guo; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022, pp. 599-608

Abstract


Shadow removal is an important yet challenging restoration task. State-of-the-art shadow removal methods usually require paired datasets for training. Existing shadow removal datasets lack large-scale quantity and scene diversity. Hence, models trained on such datasets have poor generalization ability. This paper proposes a simple yet robust shadow simulation method to simulate shadow on the grayscale. The proposed shadow simulation method can be applied to arbitrary shadow-free images and masks to generate corresponding shadow images. With our shadow simulation method, we can generate a large-scale and diverse paired shadow removal dataset. Besides, we introduce a two-stage framework, Gray-to-Color Shadow Removal Network (G2C-DeshadowNet) for shadow removal. The first stage is a Grayscale Enhancement Network, which attempts to remove shadows on the grayscale. The second stage is a Colorization Network, which attempts to colorize the grayscale shadow-free image. Extensive experiments on ISTD+, SRD, and SBU datasets show that G2C-DeshadowNet outperforms state-of-the-art methods and has better generalization ability.

Related Material


[pdf]
[bibtex]
@InProceedings{Gao_2022_CVPR, author = {Gao, Jianhao and Zheng, Quanlong and Guo, Yandong}, title = {Towards Real-World Shadow Removal With a Shadow Simulation Method and a Two-Stage Framework}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops}, month = {June}, year = {2022}, pages = {599-608} }