Eli Kaminsky, Michael Werman*

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

1 Scopus citations


Finding shadows in images is useful for many applications, such as white balance, shadow removal, or obstacle detection for autonomous vehicles. Shadow segmentation has been investigated both by classical computer vision and machine learning methods. In this paper, we propose a simple Convolutional-Neural-Net (CNN) running on a PC-GPU to semantically segment shadowed regions in an image. To this end, we generated a synthetic set of shadow objects, which we projected onto hundreds of shadow-less images in order to create a labeled training set. Furthermore, we suggest a novel loss function that can be tuned to balance runtime and accuracy. We argue that the combination of a synthetic training set, a simple CNN model, and loss function designed for semantic segmentation, are sufficient for semantic segmentation of shadows, especially in outdoor scenes.

Original languageAmerican English
Title of host publicationImage Analysis and Recognition - 15th International Conference, ICIAR 2018, Proceedings
EditorsBart ter Haar Romeny, Fakhri Karray, Aurelio Campilho
PublisherSpringer Verlag
Number of pages9
ISBN (Print)9783319929996
StatePublished - 2018
Event15th International Conference on Image Analysis and Recognition, ICIAR 2018 - Povoa de Varzim, Portugal
Duration: 27 Jun 201829 Jun 2018

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume10882 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349


Conference15th International Conference on Image Analysis and Recognition, ICIAR 2018
CityPovoa de Varzim

Bibliographical note

Publisher Copyright:
© 2018, Springer International Publishing AG, part of Springer Nature.


  • Basin-loss function
  • IoU - Intersection over Union
  • Shadow detection


Dive into the research topics of 'ShadowNet'. Together they form a unique fingerprint.

Cite this