Method and apparatus for video object segmentation

摘要

Methods and apparatus for video object segmentation are provided, suitable for use in a super-resolution system. The method comprises alignment of frames of a video sequence, pixel alignment to generate initial foreground masks using a similarity metric, consensus filtering to generate an intermediate foreground mask, and refinement of the mask using spatio-temporal information from the video sequence. In various embodiments, the similarity metric is computed using a sum of squared differences approach, a correlation, or a modified normalized correlation metric. Soft thresholding of the similarity metric is also used in one embodiment of the present principles. Weighting factors are also applied to certain critical frames in the consensus filtering stage in one embodiment using the present principles.

主权项

1. An apparatus for video object segmentation, comprising:
a memory and frame alignment mechanism that stores a plurality of frames of video and aligns one or more reference frames with a current frame containing a video object; circuitry that generates an intermediate mask for the current frame based on a neighborhood similarity metric; and a processor that refines the intermediate mask by using information from at least one video frame to generate a mask representative of the video object in the current frame: circuitry that generates initial foreground masks for a current frame with respect to each aligned reference frame based on a neighborhood similarity metric; and a generator that combines information from the initial foreground masks to generate an intermediate mask for the current frame before refining the intermediate foreground mask.