Computational Depth Sensing: Toward High Performance Commodity Depth Cameras

Depth information plays an important role in a variety of applications, including manufacturing, medical imaging, computer vision, graphics, and virtual/augmented reality (VR/AR). Depth sensing has thus attracted sustained attention from both academia and industry communities for decades. Mainstream depth cameras can be divided into three categories: stereo, time of flight (ToF), and structured light. Stereo cameras require no active illumination and can be used outdoors, but they are fragile for homogeneous surfaces. Recently, off-the-shelf light field cameras have demonstrated improved depth estimation capability with a multiview stereo configuration. ToF cameras operate at a high frame rate and fit time-critical scenarios well, but they are susceptible to noise and limited to low resolution [3]. Structured light cameras can produce high-resolution, high-accuracy depth, provided that a number of patterns are sequentially used. Due to its promising and reliable performance, the structured light approach has been widely adopted for three-dimensional (3-D) scanning purposes. However, achieving real-time depth with structured light either requires highspeed (and thus expensive) hardware or sacrifices depth resolution and accuracy by using a single pattern instead.