Towards Semantic labeling of 3D Point Clouds

                                                                             by Akshay Jain

 Introduction

   This report describes the progress done by the author in developing a system to use semantic information for understanding scenes and identifying objects. The first two steps to perform this task are segmentation and feature computation.

   For the segmentation step, the point cloud is segmented into different regions based on a region growing algorithm similar to Euclidean clustering but adding smoothness constraint with the Euclidean distance. Then for feature extraction, two sets of features are calculated to represent each segment. These are 2D features which represent the visual appearance of the segments and 3D features which describe their geometry.

  The algorithm was able to classify objects with a classification rate of 76 % when the objects under consideration were different (table top, chair back rest, monitor). It is observed that in all the cases that the classification using only visual features is better than the using only 3D features. When the number of objects to be classified are increased, the classification rate is dropped. But even for small number of objects, for example, wall and floor, the classification rate is not high. The floor is confused a lot with the wall.