Multiscale Saliency Detection for Colored 3D Point Clouds Based on Random Walk

Abstract

Saliency detection for 3D visual data has been actively studied, but relatively little effort has been made to detect both of the geometric and photometric saliency for large-scale colored 3D point clouds (LSC3DPCs). We propose a random walk based multiscale saliency detection algorithm for LSC3DPCs acquired by terrestrial light detection and ranging devices. We employ Fast Point Feature Histogram descriptor and Lab colors to estimate the geometric and photometric features of points, respectively. We partition an input LSC3DPC model into supervoxel clusters at three different scales of octree. Then we build a fully-connected graph of clusters at each scale such that an edge connecting two clusters with more dissimilar features to each other is assigned a higher weight. We perform random walk simulation on the graphs at multiple scales to yield multiscale saliency maps, respectively, which are then averaged together to generate a final saliency map. Experimental results show that the proposed method estimates the global and local saliency of LSC3DPCs more faithfully compared with the existing method.