A new video segmentation technique and evaluation

Abstract EN

This paper introduces a new segmentation technique that takes motion into consideration.

The technique works by dividing each frame into blocks of a certain size; starting from a reference frame, each block is compared with all the blocks in a region surrounding the corresponding block in the next frame.

To measure the similarity between frames, a technique combining mean square error and hit-ratio is used; the mean square error is used to measure the similarity between blocks, and the percentage (hit-ratio) of blocks passing that similarity test must exceed a preset threshold.

Different values of block sizes, motion displacement, and hit-ratios are considered in this paper to optimize segmentation.

A block size of 16, a displacement of 6 and a hit-ratio of 70% are found to be the best.

Our segmentation technique has advantages over others such as motion sensitivity, adjustable segmentation granularity, and close agreement with subjective notion of visual segments.