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 language||American English|
|Title of host publication||Image Analysis and Recognition - 15th International Conference, ICIAR 2018, Proceedings|
|Editors||Bart ter Haar Romeny, Fakhri Karray, Aurelio Campilho|
|Number of pages||9|
|State||Published - 2018|
|Event||15th International Conference on Image Analysis and Recognition, ICIAR 2018 - Povoa de Varzim, Portugal|
Duration: 27 Jun 2018 → 29 Jun 2018
|Name||Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)|
|Conference||15th International Conference on Image Analysis and Recognition, ICIAR 2018|
|City||Povoa de Varzim|
|Period||27/06/18 → 29/06/18|
Bibliographical noteFunding Information:
This research was supported by the Israel Science Foundation and by the Israel Ministry of Science and Technology.
© 2018, Springer International Publishing AG, part of Springer Nature.
- Basin-loss function
- IoU - Intersection over Union
- Shadow detection