PhD Defence | Recurrent Motion in Vision
In his thesis, Recurrent Motion in Vision, Tom Runia explores the concept of recurrent motion patterns in video. Runia completed his research under the supervision of Cees Snoek and Armold Smeulders (both from the UvA).
The thesis, consisting of two parts, begins with a study on the origin of periodic motion which results in its categorization into fundamental cases. Based on the theory, the thesis introduces novel solutions for estimating repetition in real-world videos by means of counting and measuring physical object properties related to visual recurrence such as cloth in the wind. The proposed solutions cover a wide range of aspects in computer vision, from traditional signal processing to modern day deep learning algorithms.
Recurrent motion is ubiquitous in the visual world around us. In a typical mundane scene such as having breakfast, visual rhythm can be perceived as the stirring in our coffee, the cutting of oranges, spreading marmalade on toast and the chewing motion of our jaws. Given the ubiquity of recurrent motion in the visual world, it is worthwhile to understand its origin and how it is perceived by our visual system. Apart from cognitive importance and academic curiosity, understanding the origins of recurrent motion in vision and our ability to detect, localize and count it has numerous important applications. In computer vision, the presence of recurrent motion has been leveraged to infer 3D structure, estimate depth, classify human actions, categorize sports video, calibrate cameras and to find duplicate video content. These examples underscore the value of recurrent motion in real-world problems and serve as the primary motivation for Runia’s study on the origin, appearance and value of recurrent motion in vision.
In his thesis, starting from the motion field induced by a moving object, Runia categorizes fundamental cases of intrinsic periodic motion through a decomposition of the motion. For the 2D perception of 3D periodicity as appearing in video he derives a categorization of fundamental cases of recurrent perception from the differential operators acting on the motion field. These insights are then leveraged for methods to detect, localize and count recurrent motion in video. For example, to handle cases of non-stationary repetition, Runia adopts the continuous wavelet transform to extract time-varying frequency motion from video. Furthermore, the thesis introduces methods for measuring physical object properties from real-world video such as the appearance of cloth in the wind.
Link to watch PhD defence live via Youtube will be released approximately 1 week before the defence.