[campus icon] Accesskey [ h ] University of Paderborn - Home
DE deutsch
Die Universität der Informationsgesellschaft


Semantic-aware Lightweight Visual Place Recognition System for Severe Viewpoint and Appearance Variations
Date: 2020/01/08
Time: 16:30 h
Place: P 1.4.17
Author(s): Bashar Tomeh

On Wednesday, January 8, Bashar Tomeh will present the results of his master thesis with the title:

Semantic-aware Lightweight Visual Place Recognition System for Severe Viewpoint and Appearance Variations


Visual Place Recognition has played a significant role in different computer vision and robotics tasks. It can be seen as an image retrieval task where the system has to match a query image with images in the previously created database. Such a system can be used in image searching engine, autonomous driving as well as in a loop-closure system where a mobile robot tries to compensate for the arising errors when creating a map of an unknown environment by detecting previously visited places. The challenging problem is that the images in the database and the query image can be taken under different perceptual conditions as well as different viewpoints. Besides this, the image has to be represented in a single compact vector for efficient searching. This master's thesis focuses on creating an image representation that is robust to the severe viewpoint and appearance changes while preserving a low computational cost to create a real-time system feasible for a resource-constrained mobile robot. Furthermore, the effect of including the semantic understanding of the place in the image representation will be explored in this study. In this manner, two systems are proposed using the knowledge from previous work with a vector of locally aggregated descriptors applied as a convolutional layer, where the two systems differ in using the semantic information from a segmentation mask. The overall goal is a semantic-aware lightweight convolutional neural network system, which can learn robust image representation to distinguish places in presence of high visual ambiguity and viewpoint variations. The performance of the proposed systems is evaluated on benchmark datasets, considering the precision-recall scores, computational efficiency and memory usage.

The presentation starts at 16:30h in room P 1.4.17.