Hierarchical Motion Estimation for Small Objects in Frame-Rate Up-Conversion

Abstract

Block-based hierarchical motion estimations are widely used for frame interpolation in frame-rate up-conversion and are successful in generating high-quality interpolations. However, it still fails in the motion estimation of small objects when a background region moves in a different direction. This is because the motion of small objects is neglected by the down-sampling and over-smoothing operations at the top level of image pyramids in the maximum a posterior (MAP) method. Consequently, the motion vector of small objects cannot be detected at the bottom level, and therefore, the small objects often appear deformed in an interpolated frame. This paper proposes a novel algorithm that preserves the motion vector of the small objects by adding a secondary motion vector candidate that represents the movement of the small objects. This additional candidate always propagates from the top to the bottom layers of the image pyramid. Experimental results demonstrate that the intermediate frame interpolated by the proposed algorithm significantly improves the visual quality when compared with conventional MAP-based frame interpolation.