A Robust Oversegmentation Algorithm using Colour and Geometric Cues
Amara Kiran, K. G. Sreeni · 2014
Superpixels are now widely used to enhance segmentation efficiency and computational speed for various vision related tasks such as object detection and scene understanding. Available superpixel algorithms rely mostly on colour and spatial proximity to perform oversegmentation. The problem arises when the object and the background are of same colour, causing all of these state-of-the-art methods to fail in segmenting out the object precisely. Here a novel oversegmentation method is proposed which uses depth/range data to overcome the foresaid challenge. A distance measure is formed using RGB-D information and with the help of this measure, the pixels are iteratively clustered. As depth data provides geometric details about the object, using it along with colour and spatial information provides a better segmentation accuracy. We compare our approach with 5 state-of-the-art methods and shows that the proposed method outperforms all of them.