Fast Stereo Depth Estimation in Smartphone Devices with Narrow Baseline

Abstract

With an expanding market of mobile devices consisting of dual camera configuration, depth estimation has become a core technology for various camera applications. However, due to the constraints imposed by camera configurations and relatively lower computational capabilities of mobile devices, state of the art methods, do not fare well on embedded devices. In this work, we address the challenges in designing a fast depth estimation method for asymmetric dual camera, with narrow baseline and limited computing power on a smartphone device. We propose a novel approach to efficiently compute accurate matching cost volume using sub-pixel steps. Additionally, a modified Semi Global Matching cost optimization and confidence measure based on binary edge maps are used for dense depth map estimation. To validate the proposed method, we use a dataset consisting of 600 stereo image pairs captured using two smart phones with dual cameras. The proposed method demonstrates significant visual improvements in areas involving repeating patterns, smooth regions and complex object boundaries compared to prior methods.