cuRCD: Region covariance descriptor CUDA implementation

• Published in: Multimedia tools and applications Vol. 80; no. 13; pp. 19737 - 19751
• Main Authors: Asan, M. Ali, Ozsoy, Adnan
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.05.2021; Springer Nature B.V
• Subjects: Algorithms; Computer Communication Networks; Computer Science; Covariance matrix; Data Structures and Information Theory; Feature selection; Machine learning; Multimedia; Multimedia Information Systems; Neural networks; Real time; Robustness (mathematics); Special Purpose and Application-Based Systems; CUDA; Parallel region covariance; GPGPU; Real time object detection
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-021-10644-2

Region covariance is a robust feature descriptor that allows the use of even the simplest image features like intensity and gradient combined to form a well-performing descriptor for regions on the image. Beyond its robustness, it requires many identical heavy computations on different parts of input data which makes it a good candidate for parallel execution. In this manuscript, we present a real-time parallel implementation of the region covariance which, to our best knowledge, is the first in the literature. We experimented against existing implementations and achieved 6 times faster execution time over vectorized CPU parallel implementation that provides necessary speed up for real-time processing. Additionally, we improved the existing integral image calculation method on CUDA, reducing memory usage by 50%, achieving the fastest computation speed compared to exist- ing solutions, and improved the covariance matrix comparison metric by using a distance metric that is lightweight to compute and easy to implement.