An Effective 3D Shape Descriptor for Object Recognition with RGB-D Sensors

Zhong Liu,Changchen Zhao,Xingming Wu,Weihai Chen

doi:10.3390/s17030451

Abstract

RGB-D sensors have been widely used in various areas of computer vision and graphics. A good descriptor will effectively improve the performance of operation. This article further analyzes the recognition performance of shape features extracted from multi-modality source data using RGB-D sensors. A hybrid shape descriptor is proposed as a representation of objects for recognition. We first extracted five 2D shape features from contour-based images and five 3D shape features over point cloud data to capture the global and local shape characteristics of an object. The recognition performance was tested for category recognition and instance recognition. Experimental results show that the proposed shape descriptor outperforms several common global-to-global shape descriptors and is comparable to some partial-to-global shape descriptors that achieved the best accuracies in category and instance recognition. Contribution of partial features and computational complexity were also analyzed. The results indicate that the proposed shape features are strong cues for object recognition and can be combined with other features to boost accuracy.

Highlights

In the field of computer vision, the last few decades have considered object recognition to be a fundamental task and is still an active research topic
We evaluated our descriptor at two levels: category recognition and instance recognition, which are two basic capabilities for vision-based service robots, manipulators and surveillance systems
The results show that the proposed descriptor is comparable to the shape descriptors that achieve the best accuracies both in category and instance recognition, with slightly better accuracy in category recognition and lightly lower accuracy in instance recognition

Summary

Introduction

In the field of computer vision, the last few decades have considered object recognition to be a fundamental task and is still an active research topic. Deep learning technology and convolutional neural networks have been extensively developed so that a great number of tasks in computer vision, including object recognition, have seen dramatic improvements due to the advances of deep learning. In some cases where both optic cameras and infrared cameras are available, deep learning technology may not be the best solution for object recognition due to the heavy computational load and unexpected issues when applying deep learning on both RGB and depth data [7]. With the help of RGB-D sensors, the data from multimodal sources provide much more cues for recognition than the plain RGB data. The purpose of this paper is to exploit the ability of some shape features extracted from RGB-D sensors for object recognition

Objectives

Methods

Results

Conclusion