TSD: A new dataset for shadow detection of transparent objects

Abstract

In recent years, with the rapid development of deep learning and its wide application in the field of computer vision, a series of shadow detection algorithms based on deep learning have been proposed on public datasets such as SBU and ISTD. These shadow detection algorithms verify better performance than traditional shadow detection algorithm based on the physical model. However, the existing shadow detection dataset only performs image acquisition for non-transparent objects, and ignores the requirement for shadow removal of transparent objects in practical applications. Therefore, in this article, we will focus on the production of the transparent object shadow detection dataset, and propose a method of making synthetic dataset pictures through Blender software. We propose a new dataset for shadow detection of transparent objects and the new dataset contains 800 images including 100 images with complex backgrounds, 233 images with bright backgrounds and 467 images without backgrounds. We also verify it on the existing shadow detection algorithm. The experimental results show that the datasets we made has good characteristics and can be better used for the training of shadow detection networks and the detection of shadow positions of transparent objects.