Comparing Fusion Methods for 3D Object Detection

Abstract

One of the main problems in computer vision that leads digital technologies to transform our business and social life is object detection. The solutions to the problem have various application areas such as security systems, surveillance, shopping applications, and much more. Significant performance gain have been achieved by using the popular deep learning methods that are applied on 2D RGB images. With the availability of low-cost 3D sensors, methods that incorporates 3D data with 2D RGB data using existing deep network architectures became more popular to increase performance further. In this work, different data level and feature level fusion strategies have been analyzed to incorporate 3D depth and 2D RGB data with existing architectures to assess their effects on the performance. These methods were tested on the real RGB-D benchmark datasets available in the literature, and the accuracy results were compared to each other in their object types of group. © 2021 Elsevier B.V., All rights reserved.