Experimental implementations of stereo matching algorithms in Halide

Abstract

Currently, most stereo matching algorithms focus their efforts on increasing accuracy at the price of losing run-time performance. However, applications such as robotics require high performance stereo algorithms to perform real time tasks. The problem is due to the difficulty of hand optimizing the complicated stereo matching pipelines. Halide is a programming language that has been widely used in writing high-performance image processing codes. In this work, we explore the usability of Halide in the area of real-time stereo algorithms by implementing several stereo algorithms in Halide. Because of Halide's ability to reduce the computation cost of dense algorithms, we focus on local dense stereo matching algorithms, including the simple box matching algorithm and the adaptive window stereo matching algorithms. Although we have found Halide's limitation in scheduling dynamic programming and recursive filters, our results demonstrate that Halide programs can achieve comparable performance as hand-tuned programs with much simpler and understandable code. Lastly, we also include a design solution to support dynamic programming in Halide.