3D spatial pyramid: descriptors generation from point clouds for indoor scene classification

Abstract

Traditionally, the indoor scene classification problem has been approached from a 2D image recognition point of view. In most visual scene classification systems, a descriptor for the input image is generated to obtain a suitable representation that includes features related to color, shape or spatial information. Techniques based on the use of a spatial pyramid have proven to be adequate to perform this step. In the past years, on the other hand, 3D sensors have become widely available, which allows to include new information sources to the framework previously described. In this work we rely on RGB-D data to extend the spatial pyramid approach, aimed at building descriptors that can lead to a more robust representation against changing lighting conditions. The proposed descriptors are evaluated on the RobotVision@ImageCLEF-2013 benchmark dataset, remarkably outperforming state-of-the-art 3D local and global descriptors.