Optimized Code Generation of Computer Vision Transformations using Sparse Matrices

Modern computer vision image transformations can benefit from sparse input data in complex pipelines. However, such routines are difficult to implement efficiently with sparse matrices.Therefore, we propose a domain-specific compiler that computer vision specialists can use. The tool reads a mathematical specification describing a transformation, relating the input and output image pixels, together with a generic C code template, and generates from them an efficient C implementation using sparse matrices. The compiler performs a novel code optimization to increase performance by traversing inorder the input sparse image. For this, our DSL compiler needs to perform in certain cases symbolic mathematical optimization on the equations obtained from the specification.We use our DSL compiler to help us develop the sparse version of useful transformations from a computer vision pipeline performing image alignment. We achieve an average sequential performance increase of 5.34 times on an x86 processor at 2.66 GHz when using sparse matrices at a sparsity level of 95.22% than when using high-definition, dense images.