Computer Vision Talks : Scale-Equivariant Siamese Tracking
Paper Abstract:
In this talk, we start by discussing the role of invariance and equivariance in computer vision tasks. We demonstrate how to build general convolutional networks equivariant to scale transformations by design. We propose the theory for scale-equivariant Siamese trackers and provide a simple recipe for making a wide range of existing trackers scale-equivariant. We demonstrate the advantage of scale-equivariant Siamese tracking over conventional models.
Speaker Bio:
Ivan Sosnovik is a Ph.D. student at UvA-Bosch Delta Lab at the University of Amsterdam. Ivan received a Master of Science at Moscow PhysTech with honors in 2017. Main research interests are symmetry learning in computer vision, learning invariances in neural networks, CNNs with a flavor of differential geometry, invertible models.
Artem Moskalev is a Ph.D. student at UvA-Bosch Delta Lab at the University of Amsterdam. Artem received an MSc in Applied Math at Skolkovo Institute of Science and Technology in 2019. Research agenda: object tracking, differentiable and combinatorial optimization, equivariance, and symmetries in computer vision.
