Spatio-temporal tube segmentation through a video metrics-based patch similarity measure

This paper presents a new method for video simplification which identifies moving or static objects in a video by grouping together all the pixels corresponding to homogeneous texture and temporal coherent spatio-temporal regions, the so-called tubes. This is achieved through the proposal of video metrics defined on the video domain, which provide patches that intrinsically and automatically adapt its spatio-temporal shape and size, and a patch-based comparison measure that is able to capture the scene similarities. Building upon this video metrics-based patch similarity measure we propose a video simplification method that results in tubes made of the pixels with the same vicinity texture content regardless camera point of view or object position changes. We present experiments analyzing the performance of the proposed approach.