Retaining descriptive power of 3D feature descriptors with much less

On Creating Low Dimensional 3D Feature Descriptors with PCA

The availability of commodity depth sensors and depth sensing capabilities on mobile devices sees the need for memory efficient and computationally cheap applications. 3D feature descriptors are extensively used in various 3D perception applications such as 3D object recognition, retrieval and SLAM. However, existing state of the art 3D feature descriptors are very high dimensional and hence require higher computational and memory resources. In this paper, Sai Manoj Prakhya and his team from A*STAR Institute for Infocomm Research (I 2R)’s Autonomous System Department try to address the problem of dimensionality reduction of 3D feature descriptors, which can have practical significance in terms of memory and computational power when deployed in real world 3D perception applications.

In order to reduce these computational and memory requirements, the team proposed to create low dimensional 3D feature descriptors using Principal Component Analysis (PCA) by learning the transformation matrices from lots of 3D feature descriptors. A dataset of a million 3D feature descriptors were created and the learnt projection matrices were made publicly available thereafter.

Three state of the art 3D feature descriptors namely SHOT, RoPS and FPFH were selected and systematically applied with the proposed method to create their low dimensional variants. Experimental results have shown that PCA-SHOT with 50 dimensions, PCA-RoPS with 30 dimensions and PCA-FPFH with 15 dimensions can achieve more than 90% of RRR performance, when compared to their parent descriptors, which require 352 dimensions (SHOT), 135 dimensions (RoPS) and 33 dimensions (FPFH), respectively. Their performances were clearly highlighted with respect to the dimensionality while benchmarked against the full dimensional ones. This thorough analysis can aid the end user to directly choose the low dimensional variant that is suitable for their purpose based on the target performance requirements.

Sai Manoj Prakhya highlighted that the main stream research on 3D feature descriptors until now has always been on creating more efficient ones that offer high performance. The problem of dimensionality and computational requirements has not been dealt in the main stream in 3D domain yet. Hence, his team has presented a benchmark for low dimensional 3D feature descriptors by presenting their performance.

* This paper clinched the Best Paper Award in IEEE Region Ten Conference (TENCON) 2017.