A new MRF model for robust estimate of occlusion and motion vector fields

This paper proposes a new Markov random field (MRF) model for the detection of occluded regions in image sequences. Motion vectors are not defined in an occluded region, thus the regions with high motion compensated prediction error are commonly regarded as occluded regions. However, badly motion compensated pixels will also appear as occluded pixels, making it difficult to distinguish the true occluded pixels from the poorly motion compensated regions. The proposed MRF model addresses this problem by incorporating motion information into the occlusion model. This is derived from the observation that occlusion occurs when objects move. It is found that an accurate occlusion region can be detected and better motion estimation can be made with the new model using iterated conditional modes (ICM)