Robots move and gather information sequentially; this additional information is often used to improve Visual Place Recognition (VPR) and localization ability of mobile robots. However, these methods rely on strong assumptions of repeat traverses in the ‘same sequential order’, limiting their applicability in general. This PhD project will explore novel methods, beyond bidirectional recurrent networks or fully-order-unaware aggregation, to develop sequence-based place representations and sequence-based place matching.
The student will get an opportunity to delve into several deep learning-based techniques, especially attention-based mechanisms composed of graph convolutions or transformer-style architectures.