MRF-MAP-MFT visual object segmentation based on motion boundary field

摘要

In our earlier work, a two-pass motion estimation algorithm (TPA) was developed to estimate a motion field for two adjacent frames in an image sequence where contextual constraints are handled by several Markov random fields (MRFs) and the maximum a posteriori (MAP) configuration is taken to be the resulting motion field. In order to provide a trade-off between efficiency and effectiveness, the mean field theory (MFT) was selected to carry out the optimization process to locate the MAP with desirable performance. Given that currently in the disciplines of digital library [IEEE Trans. PAMI 18 (8) (1996); IEEE Trans. Image Process. 11 (8) (2002) 912] and video processing [IEEE Trans. Circ. Sys. Video Tech. 7 (1) (1997)] of utmost interest are the extraction and representation of visual objects, instead of estimating motion field, in this paper we focus on segmenting out visual objects based on spatial and temporal properties present in two contiguous frames in the same MRF-MAP-MFT framework. To achieve object segmentation, a "motion boundary field" is introduced which can turn off interactions between different object regions and in the mean time remove spurious object boundaries. Furthermore, in light of the generally smooth and slow velocities in-between two contiguous frames, we discover that in the process of calculating matching blocks, assigning different weights to different locations can result in better object segmentation. Experimental results conducted on both synthetic and real-world videos demonstrate encouraging performance.