3:00 pm to 4:00 pm
GHC 6501
Abstract: The majority of existing vision-based methods perform multi-object tracking in the image domain. Yet, in mobile robotics and autonomous driving scenarios, pixel-precise object localization and trajectory estimation in 3D space are of fundamental importance. Furthermore, the leading paradigms for vision-based multi-object tracking and trajectory prediction heavily rely on object detectors and effectively limit tracking and motion prediction to a set of predefined classes, while the set of object classes that appear in the real-world is unbounded.
In the first part of my talk, I will present our recent work on Category-Agnostic Multi-Object Tracking (CAMOT) and 4D Generic Video Tubes (4D-GVT). Both leverage recent developments in learning-based object proposal generation to estimate trajectories of arbitrary objects from the stereo video and achieve remarkable generalization to unseen object classes.
In the second part, I will talk about our work on joint Multi-Object Tracking and Segmentation (MOTS), for which we create dense pixel-level annotations for two existing tracking datasets and we propose a new method which tackles detection, tracking, and segmentation jointly with a single convolutional network.
I will conclude my talk with our recent results on learning to estimate the precise 3D relative motion of objects using LiDAR/RGB-D sensors.
Bio: Aljosa Osep is at the final stage of his doctoral studies at RWTH Aachen University (his defense date is in June 2019) under the supervision of Prof. Dr. Bastian Leibe and is joining the Dynamic Vision Group at the Technical University in Munich starting in August 2019. He obtained MSc degree in Computer Science at the University of Bonn in 2013 with a focus on multi-view 3D reconstruction of specular objects. His main research interests are on the intersection of Computer Vision, Mobile Robotics and Machine Learning. During his Ph.D., he was mainly working on generic multi-object tracking, segmentation, and learning-in-the-wild from unlabeled video in mobile robotics scenarios.