TY - GEN
T1 - MGRLN-Net
T2 - 16th Asian Conference on Computer Vision, ACCV 2022
AU - Jie, Leiping
AU - Zhang, Hui
N1 - This work was supported by the National Natural Science Foundation of China (62076029), Guangdong Science and Technology Department (2017A030313362), Guangdong Key Lab of AI and Multi-modal Data Processing (2020KSYS007). and internal funds of the United International College (R202012, R201802, R5201904, UICR0400025-21).
Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023.
PY - 2023/3/2
Y1 - 2023/3/2
N2 - Although significant progress has been made in single-image shadow
detection or single-image shadow removal, only few works consider these
two problems together. However, the two problems are complementary and
can benefit from each other. In this work, we propose a Mask-Guided
Residual Learning Network (MGRLN-Net) that jointly estimates shadow mask
and shadow-free image. In particular, MGRLN-Net first generates a
shadow mask, then utilizes a feature reassembling module to align the
features from the shadow detection module to the shadow removal module.
Finally, we leverage the learned shadow mask as guidance to generate a
shadow-free image. We formulate shadow removal as a masked residual
learning problem of the original shadow image. In this way, the learned
shadow mask is used as guidance to produce better transitions in
penumbra regions. Extensive experiments on ISTD, ISTD+, and SRD
benchmark datasets demonstrate that our method outperforms current
state-of-the-art approaches on both shadow detection and shadow removal
tasks.
AB - Although significant progress has been made in single-image shadow
detection or single-image shadow removal, only few works consider these
two problems together. However, the two problems are complementary and
can benefit from each other. In this work, we propose a Mask-Guided
Residual Learning Network (MGRLN-Net) that jointly estimates shadow mask
and shadow-free image. In particular, MGRLN-Net first generates a
shadow mask, then utilizes a feature reassembling module to align the
features from the shadow detection module to the shadow removal module.
Finally, we leverage the learned shadow mask as guidance to generate a
shadow-free image. We formulate shadow removal as a masked residual
learning problem of the original shadow image. In this way, the learned
shadow mask is used as guidance to produce better transitions in
penumbra regions. Extensive experiments on ISTD, ISTD+, and SRD
benchmark datasets demonstrate that our method outperforms current
state-of-the-art approaches on both shadow detection and shadow removal
tasks.
KW - Shadow detection and removal
KW - Multi-task learning
KW - Masked residual learning
UR - http://www.scopus.com/inward/record.url?scp=85151055359&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-26313-2_28
DO - 10.1007/978-3-031-26313-2_28
M3 - Conference proceeding
AN - SCOPUS:85151055359
SN - 9783031263125
T3 - Lecture Notes in Computer Science
SP - 460
EP - 476
BT - Computer Vision – ACCV 2022
A2 - Wang, Lei
A2 - Gall, Juergen
A2 - Chin, Tat-Jun
A2 - Sato, Imari
A2 - Chellappa, Rama
PB - Springer
CY - Cham
Y2 - 4 December 2022 through 8 December 2022
ER -