Invited talks

Tomas Pajdla 

Algebraic Vision

The geometry of computer vision was, for a long time, building on linear algebra and projective geometry. Since about 2002, non-linear algebraic methods started showing their strength. Algebraic geometry techniques were used to understand old geometrical problems better and construct new efficient solvers for computing camera geometry within the RANSAC estimation paradigm. In this talk, I will present some new results and sketch the algebraic techniques that led to new interesting results in computer vision. 

Andrew Fitzgibbon

AI Hardware and Real-World AI

AI is fast becoming a significant consumer of the world’s computational power, so it is crucial to use that power wisely and efficiently. Our approaches to doing so must span all levels of the research stack: from fundamental theoretical understanding of the loss surfaces and regularization properties of machine learning models, to efficient layout at the transistor level of floating-point multipliers and RAM. I will talk about projects, such as real-time computer vision on the Microsoft HoloLens HPU (about 3.5 GFLOPS), which required extreme efficiency in both objective and gradient computations, and how this relates to the training of massive AI models on Graphcore’s IPU (about 350 TFLOPS). Key to this work is how we empower programmers to communicate effectively with such hardware, and how we design frameworks and languages to ensure we can put theory into practice. So this talk contains aspects of: mathematical optimization, automatic differentiation, programming languages, and silicon design. Despite this range of topics, the plan is for it to be accessible and useful to anyone who loves computers.

Danijel Skočaj

Data-driven learning-based surface anomaly detection

In the last decade, computer vision has made tremendous progress, driven by rejuvenated deep learning and fuelled by large quantities of data that have become readily available. In the last couple of years, the data-driven learning-based approach has also started entering a more conservative engineering discipline of machine vision. It has proven to be a very promising alternative to the main development paradigm, which is based on developing hand-engineered specific solutions for machine vision problems. The learning-based approach facilitates more general, efficient, flexible and economical development, deployment and maintenance of machine vision systems. In this talk, we will discuss this new development paradigm. Several data-driven approaches to surface defect detection will be presented, ranging from unsupervised to fully supervised methods. We will discuss the advantages of these approaches and the challenges they face and address the role and opportunities of learning-based approaches for efficient visual inspection in the framework of the Industry 4.0 paradigm.