Title of Seminar:  Learning from Motion

Speaker Institution:  Inria, France

Speaker’s title:  PhD Student

Speaker Homepage:  http://thoth.inrialpes.fr/people/tokmakov/

BIO:  Pavel Tokmakov is a PhD student at the THOTH team at Inria, France under supervision of Karteek Alahari and Cordelia Schmid. In his PhD he works on studying the role of motion in object recognition. He obtained his Master’s degree at the University of Bonn, Germany, where he worked on Statistical Relational Learning and Interactive Knowledge Discovery. His research interests also include structured and self-supervised learning.

Abstract:  Weakly-supervised learning for semantic segmentation has received a lot of attention since the introduction of fully-convolutional networks (FCNs). Despite this fact, the performance of image-based methods remained significantly behind the fully-supervised approaches. One of the main issues is the inability of weakly-supervised methods to accurately capture boundaries of the objects. In our work [Tokmakov et al., ECCV’16] we propose to utilize motion cues in videos to obtain constraints on object shapes for free. We then integrate motion cues with the semantic model being trained to improve the labels as learning progresses. Inspired by the utility of motion we turn to learning-based methods for moving object segmentation in videos and propose the first CNN for motion segmentation [Tokmakov et al., CVPR’17], which is later extended to a video object segmentation framework by augmenting the motion network with an appearance stream and a visual memory module [Tokmakov et al., ICCV’17].