Tracking the Known and the Unknown by Leveraging Semantic Information

Abstract

Current research in visual tracking is largely focused on the generic case, where no prior knowledge about the target object is assumed. However, many real-world tracking applications stem from specific scenarios where the class or type of object is known. In this work, we propose a tracking framework that can exploit this semantic information, without sacrificing the generic nature of the tracker. In addition to the target-specific appearance, we model the class of the object through a semantic module that provides complementary class-specific predictions. By further integrating a semantic classification module, we can utilize the learned class-specific models even if the target class is unknown. Our unified tracking architecture is trained end-to-end on large scale tracking datasets by exploiting the available semantic metadata. Comprehensive experiments are performed on five tracking benchmarks. Our approach achieves state-of-the-art performance while operating at real-time frame-rates.

Publication
In British Machine Vision Conference, BMVC 2019.
Ardhendu Tripathi
Ardhendu Tripathi
PhD Student, ETH Zurich
Martin Danelljan
Martin Danelljan
Researcher

Researcher in Computer Vision and Machine Learning at ETH Zurich

Related