Saliency Detection in Large Point Sets – Supplementary 1 Quantitative evaluation

We evaluated our algorithm on the Benchmark for 3D Interest Point Detection Algorithms [2], whose goal is to provide quantitative evaluation for detecting interest points vs. human-marked points. The benchmark contains 43 models organized in two data sets. Data-set A includes 24 of the models for which 23 human subjects have marked the interest points. Dataset B contains all the 43 models marked by at least 16 human subjects. The benchmark evaluates six techniques: Mesh saliency [4], salient points defined by Castellani et al. [1], 3D-Harris [6], 3D-SIFT [3], scale-dependent corners [5], and Heat Kernel Signature (HKS) [7].