Linear Weighted Median Filtering for Stereo Disparity Refinement

Abstract

To solve contradictions between the error disparities removal and the edges preserving in the disparity refinement step, a novel linear weighted median filter is proposed and applied to the disparity refinement for stereo algorithms. First, the classic average aggregation strategy in the guided image filter (GIF) is replaced by a weighted aggregation based on the mean of square error (MSE), and a novel edge-preserving filtering named WAGIF is proposed. The WAGIF achieves filtered images with sharper edges than those via GIF by applying weighted aggregation. Then, the proposed WAGIF is applied to design a weighted median filtering. With the assistant of the local histogram technology, the proposed weighted median filtering has a linear computational complexity. Furthermore, the experiments show that disparity errors and holes in disparity maps are removed significantly by the proposed refinement approach while edges are preserved well. The accuracy of the final refined disparity maps is improved significantly, even the proposed approach is applied to some classic stereo algorithms.