A dense depth estimation method using superpixels

Conventional stereo matching or depth estimation algorithms always provide incomplete disparity map. These pixels without depth estimation in the map are named depth gaps. Weak texture and occluded areas are main source of depth gaps. We propose a novel method to assign good depth estimation on the areas above. Our algorithm combines state-of-art superpixel segmentation approach and linear filter. First we do superpixel segmentation on reference image, after this every pixel has a label determines which superpixel it belongs to. Then merging distance between superpixels in spatial space and color space, we apply linear filter on every superpixel. We evaluate the performance of our algorithm with some classic stereo datasets to show the promotion we obtained.