Design of Class in Unknown Object Segmentation Focusing on 3D Object Detection in Depth Image

Abstract

We aim to improve unknown object detection. We also deal with problem of designing the optimal class for semantic segmentation using depth image. There was a problem that unknown classes of obstacles were mistaken for road in semantic segmentation using depth image. Therefore, we focus on the superiority of 3D object detection in a depth image. The depth image is good at separating between horizontal plane and 3D objects. For this reason, we develop a method for changing the number of training classes from baseline 12 classes to new 3 classes (void, plane, 3D object) for segmentation, which are optimal to detect unknown object by using depth images. As a result, IoU of unknown obstacle improve +6.9point than baseline method.