Video object segmentation by extended recursive-shortest-spanning-tree method

A new video object segmentation algorithm, which utilizes an extension of recursive shortest spanning tree (RSST) algorithm, is introduced. A 2-D affine motion model is assumed, and correspondingly, for each region, the planar approximation for the given dense motion vector field is computed. Starting from each 2×2 block as a distinct region, the algorithm recursively searches for the best pair of adjacent regions to merge. The “best pair” is defined as the one merging of which causes the least degradation in the performance of the piecewise planar motion vector field approximation. The RSST method is fast, parameter-free and requires no initial guess, unlike the existing algorithms. Moreover it is a hierarchical scheme, giving various segmentation masks from coarsest to finest. The algorithm successfully captures 3-D planar objects in the scene with acceptable accuracy in the boundaries, which can be further improved by utilizing the spatial information. Improvement over the existing European COST 211 Analysis Model (AM) is observed when the motion segmentation submodule of AM is replaced by the proposed