23.2 A 1920×1080 30fps 611 mW five-view depth-estimation processor for light-field applications

Abstract

Depth information has become essential in emerging computer vision applications. Although active sensing methods can provide an accurate indoor depth map, they have limited resolution and consume significant power, such as the 2.1W time-of-flight sensor in [1]. In contrast, depth estimation for stereo RGB images can provide high-resolution depth maps, even in outdoor or low-power scenarios. And, the depth accuracy can be increased by using multi-view light-field images. This paper presents an integrated circuit which estimates full HD (1920×1080) depth maps at 30fps and provides a tradeoff between depth accuracy and power consumption based on two-/three-/five-view stereo images. It addresses design challenges with three primary contributions: 1) a stripe buffering scheme which is designed to reduce the DRAM bandwidth induced by multi-view image access; 2) a four-bank interleaving architecture, which boosts computation performance by parallelizing belief-propagation (BP) operations; and, 3) an adaptive view selection unit, which realizes the accuracy-power tradeoff.