Depth information propagation direction algorithm for regularization

Abstract

3D camera technology is attractive subject in the computer vision area for a long time. Image based method 3D cameras have a disadvantage that depth cannot calculate correctly in plain area. We estimate the depth value at plain area with sparse depthmap to get full depthmap having the depth value for all pixels. Many researches have been studied for filling the plain region, but conventional regularization methods have problems such as mixing depth values unnecessarily and calculating inaccurate depth values near the edge. And we need fast and concise regularization algorithm for real-time operation. But, conventional regularization methods require a lot of memory and time consuming. So, we propose the new regularization method depth information propagation direction called as DPD to decrease the depth inaccuracy. And we also propose the modified version of the DPD algorithm to reduce the time consuming and satisfy the demand of real-time operation. However, DPD has limitations that cannot cover overall image and heavily depends on edge connectivity. So, we propose additional process such as box filtering, user interaction, and combination with conventional method. By applying the advanced DPD algorithm, we get better full depth map with preserving exact depth data. And it can be accelerated easily because all of advanced DPD process is designed based on scanline method. But, we need another matting algorithm for filling unfilled region. So, we should modify the conventional algorithm or develop scanline based regularization algorithm for broad region. It remains as future work.