Real-time depth estimation and view interpolation using Quasar

Abstract

In this paper, we first present a new programming framework, Quasar, for high-level programming on heterogeneous CPU and single/multi-GPU systems. Quasar consists of a high-level language, a corresponding integrated development environment (IDE), a compiler, and a run-time system. Its aim is to relieve the programmer from hardware-related implementation issues that commonly occur in CPU/GPU programming. Examples include selecting memory types, memory management, concurrent kernel/function execution, hardware texturing units, etc. In doing so Quasar allows the programmer to focus on designing, testing and improving the image/video processing algorithms themselves, rather than their implementation. A practical application of this programming framework is presented: a new real-time multi-camera depth estimation algorithm and subsequent view interpolation. Currently, for view interpolation, we obtain a processing speed of 30 frames/sec for HD video based on two views with Quasar on an NVIDIA Geforce GTX 770 GPU.

Publication
IS&T Electronic Imaging