Understanding and Enhancing Visual Place Recognition through Embedding Space Interpretability and Uncertainty Estimation

Visual Place Recognition (VPR) involves determining the geographic location of a photo based solely on its visual content. Recent advancements in Deep Learning (DL) have enabled the representation of images in high-dimensional spaces, where photos taken in the same location tend to cluster together, while images from different places are spread apart. This spatial organization makes it easier to predict locations by performing similarity searches against a database of known places. However, a key gap in current research is understanding the specific information retained in these image embeddings that allows for effective and reliable location prediction. Additionally, existing State-of-the-Art (SOTA) deterministic methods in VPR are unable to quantify the uncertainty of their predictions. This is particularly problematic in safety-critical applications, such as autonomous driving, where knowing the confidence level in a system's decision is vital for ensuring safety. This thesis addresses two main challenges: first, understanding and visualizing the essential information encoded in image embeddings, and second, providing uncertainty estimates for VPR models through post-hoc techniques. To overcome these challenges, the thesis employs Generative Artificial Intelligence models, particularly Latent Diffusion Models, to explore and visualize the content within image embeddings. Additionally, uncertainty estimation methods are incorporated to enhance the robustness and reliability of VPR systems. The contributions of this thesis provide valuable insights into the interpretability and reliability of VPR systems, offering a framework for analyzing the output of these models and incorporating uncertainty quantification during inference.