Feature extraction for efficient image and video segmentation

Color/Texture segmentation, Motion segmentation, RGB-D/T segmentation

The segmentation of sensory data of various domains is often crucial pre-processing step in many computer vision methods and applications. In this work, we propose a method that leverages the quantization of local features distributions for the depth and the temporal information. Three variants of the segmentation method is designed and evaluated reflecting various data domains: space (color and texture), temporal (motion) and depth domain. Each variant was tested on appropriate dataset showing the usability of designed method for applications like areal-image analysis, hand detection and moving-people detection. The pilot experiments shows the characteristics of the approach and computational costs of designed variants.