Super Pixel Clustering via Kernel Density Estimation

Abstract

Super pixels, which are a result of over-segmentation provide a reasonable compromise between working at pixel level versus with few optimally segmented regions. One fundamental challenge is that of defining the search space for merging. A naive approach of performing iterative clustering on the local neighborhood would be prone to under segmentation. In this paper, we develop a framework for generating non-compact super pixels by performing clustering on compact super pixels. We define the optimal search space by generating both over-segmented and under-segmented clustering of compact super pixels. Using this spatial information of the under-segmented scale, we look to improve the over-segmented scale. Our work is based on performing Kernel Density Estimation in 1D and further refining it using angular quantization. In all we propose three angular quantization formulations to generate the three scales of segmentation. Our results and comparison with the state-of-the-art super pixel algorithms show that merging a large number of super pixels with our algorithm is able to provide better results than using the underlying super pixel algorithm to obtain a smaller number of super pixels.