Edge-Based Algorithm for Multi-view Depth Map Generation

Abstract

Normalized Cross-Correlation (NCC) is a common matching measure which is insensitive to radiometric differences between stereo images. However, traditional rectangle-based NCC tends to expand the depth discontinuities. An efficient edge-based algorithm with NCC for multi-view depth map generation is proposed in this paper, which preserves depth discontinuity while remaining the advantage of robustness to radiometric differences. In addition, all pixels of initial result are classified into uncover, occlusion, reliable and unreliable by exploiting Left-Right Consistency (LRC) constraint and sequential consistency constraint. Since voting scheme will lead to errors when match windows are lack of reliable information and joint-trilateral filter will blur the depth map if employing fixed window size, especially in depth discontinuities, we combine voting scheme and joint-trilateral filter to get a better result. The experimental results show that our method achieves competitively performance.