Comparison of Supervised Learning Techniques
used for 3D Object Recognition

by Isma Hadji

Introduction

Object recognition is a hot topic in the field of computer vision and researchers in the field proposed a multitude of approaches trying to solve this problem. Any object recognition algorithm relies on three main components. A repeatable keypoint detector, a stable feature descriptor and a robust pattern recognition algorithm. In fact, the choice of the right classifier is of paramount importance and might improve or worsen the recognition results. For this specific purpose, we are proposing in this work to study the effect of the classifier chosen on the recognition rate and what makes an approach work better than another for this specific task. The classifiers evaluated are
Random Forests, Support Vector Machine, Deep Belief Networks and Low Density Separation. Our results demonstrate that with the right parameters set these classifiers results can achieve very similar accuracy.

All tests have been performed using cloud of points generated from the depth images of the well known RGB_D dataset. The feature vector used to describe the different objects is the VFH descriptor.