Deep Visual Place Recognition

ELLIS Talk by Seyran Khademi

07 February 2023 16:00 - Location: Building 28 | 0.E420 Turing | Van Mourik Broekmanweg 6 - By: ELLIS Unit Delft | Add to my calendar

Abstract:

Visual place recognition (VPR), is the process of determining the location of a single image with applications in robotics, consumer photography, social media, and archival repositories. In VPR, the most similar image to the query is retrieved from a collection of geotagged images to answer the question “where was this photo taken?”. This is a content-based image retrieval problem, where effective image representation is the key. Deep convolutional neural networks (CNNs) are proven as powerful tools for visual representation learning. This presentation will focus on using VPR for analyzing the built environment and urban scenes and will introduce the AmsterTime dataset as a benchmark for VPR in cases of severe visual domain mismatch. In this context, self-supervision will be discussed as an efficient approach to visual representation learning and a potential replacement for costly human supervision.