Robust and Efficient Visual Matching
In a majority of computer vision applications, establishing visual correspondence is a fundamental computational stage, since it provides a geometrical understanding of the data, that is crucial for high level decision making. It includes a wide range of alignment, registration or matching tasks, between color images, range scans, 3D models and more, which can provide information like the calibration of a camera, its position in space and the 3D structure and motion of objects in a scene. In such tasks, real world conditions, due to complex object or camera motion, scene geometry and illumination, are difficult to model and hence algorithms are typically challenged with high levels of noise and outlier data.
In this talk, for a range of matching problems, I will present efforts in designing rich models along with robust and efficient optimization algorithms that extend the performance limits of existing methods. I will present a very general and efficient template matcher as well as some new approaches to robust estimation that excel in the regime of high noise and outlier rates. Lastly, I will discuss the state of visual matching in the era of deep learning, including some current and future research directions.