Visual place recognition under image corruptions
Type: Bachelor EE/CS/HMI
If you are interested please contact :
Visual place recognition has received large interest from researchers in computer vision, machine learning and information retrieval. It consists of, given a query image, seeking an image depicting a similar scene in a map or database. Possible instances of this problem are the retrieval of an image containing a specific distinctive landmark (e.g. monument recognition) or the recognition of a previously visited place for robot navigation or autonomous driving. State-of-the-art algorithms for place recognition deploy convolutional neural networks, which suffer from lack of robustness when the input images are subject to corruptions as shown in  on the task of image classification.
The objective of the project is to study the robustness of convolutional networks for place recognition, with special focus on siamese architectures , when the input images are subject to common corruptions and perturbations in computer vision (e.g. noise, motion blur, zoom blur, jpeg compression, contrast, etc.).