Harris Gasparakis, OpenCV Manager at AMD, presents the "Understanding Adaptive Machine Learning Vision Algorithms and Implementing them on GPUs and Heterogeneous Platforms" tutorial at the May 2015 Embedded Vision Summit.
Machine learning algorithms are pervasive in computer vision: from object detection to object tracking to full scene recognition, generative or discriminative learning dominates the space, as it is much easier (and closer to biological systems) to program learning algorithms and learn by example, rather than directly create a program that would perform the same tasks from (yet unknown) first principles. However, unlike biological systems, machine learning algorithms tend to view learning as a pre-processing step (training), rather as an online process.
In this presentation, Harris examines archetypes of algorithms that contain magic numbers and/or fixed logic, and investigates adaptive generalizations and their implementation. Algorithms discussed include constrained energy minimization, adaptive basis function models, mixture models and graph models, with applications in object detection and tracking. Harris shows that OpenCL 2.0/HSA (Heterogeneous System Architecture) and integrated GPUs enable new design patterns and algorithms, enhancing the arsenal of tools of high performance vision scientists. When possible, Harris uses OpenCV 3.0 as the basis of his implementations.