Solving inverse problems for the analysis of fast moving objects
Project leader:
Šroubek
Others:
J. Kotera, M. Šorel, B. Zitová
Supported by:
GA18-05360S
Jointly with the:
Jiří Matas, CMP, FEL, ČVUT
Duration:
2018 - 2020
Abstract:
Objects moving fast with respect to the camera appear blurred when observed. Surprisingly this common phenomenon has not yet been considered and analyzed by the computer vision community. It is the blur that encodes information about the object motion properties. Instead of considering blur as a nuisance, the project proposes to take it as a cue for detection and tracking of fast moving objects. We formulate accurate and tractable formation models and find solutions of corresponding inverse problems which are more complex than standard blind deconvolution, since occlusion is present and blur is space variant. The complexity of the inverse problems is addressed by deriving additional priors from tracking the objects in videos and constraining the admissible set of motion blurs with the help of convolutional neural networks and blur invariants. The resulting methodology of tracking with deblurring will allow us to implement novel video analysis algorithms such as temporal superresolution, or visualization of angular velocity and deformations that are otherwise imperceptible.