Fast depth estimation using non-iterative local optimization for super multi-view images
Takanori Senoh, Koki Wakunami, Hisayuki Sasaki, Ryutaro Oi, Kenji Yamamoto · 2015
This paper proposes a high-quality fast depth estimation method based on a non-iterative edge-adaptive local cost optimization for super multi-view images (SMV). Depth candidates are increasingly updated by evaluating a cost function involving three-view matching error and depth continuity terms. A simple differentiator detects a texture edge and controls the depth continuity weight. Experimental results show an increase in speed of approximately 9.7 to 67 times with the same quality for computer graphic images and approximately 111 times with less than 1dB quality degradation for camera images compared to a graph-cuts based algorithm. Its predictable estimation time is an advantage for real-time free-navigation systems using SMV.