A new trajectory based motion segmentation benchmark dataset (UdG-MS15)

Abstract

Motion segmentation (MS) is an essential step in video analysis. Its quantitative and qualitative evaluation is largely dependent on the dataset used for testing. Although there are publicly available datasets such as Hopkins and FBMS, they have limitations in terms of number of motions, partial/complete occlusion, stopping motion, sequence length, and real life natural sequences. Due to these limitations, many recent proposals have reached nearly zero misclassification, especially for Hopkins, which leaves no room for quantitatively differentiating among proposals. In this paper, we present a new challenging trajectory based MS dataset of 15 sequences, where number of motions and sequence length have been largely increased as compared to the state of the art. An effort has been made to include all forms of distortions that are present in real life scenes. As a starting point, a preliminary benchmark evaluation using a recent and well known state of the art algorithm has been provided for this dataset.