Pose-Aware Visual Place Recognition (PAVPR)

Objective

Camera motion can be estimated from images with millimeter-level accuracy through local features, but only in carefully controlled data. In real-world conditions, performance often deteriorates under moderate illumination or viewpoint changes, dynamic scene elements or weather conditions. By contrast, Visual Place Recognition (VPR) exhibits strong robustness to visual appearance changes, yet remains limited to coarse geometric models, as it encodes image similarities without explicitly exploiting 3D information. The goal of our Pose-Aware Visual Place Recognition (PAVPR) is to advance VPR and camera motion estimation by integrating explicit 3D modeling in VPR embeddings, thereby combining robustness to appearance changes with accurate motion estimation. Building on this foundation, PAVPR will also develop localization and mapping methods that rely on holistic image descriptors rather than traditional local feature-based pipelines. This approach offers several advantages, including improved resilience to visual variations, stronger privacy guarantees, and reduced computational costs. Finally, PAVPR will design novel benchmarks that account for the geometric proportionality of VPR embeddings, addressing a key limitation of current datasets and recall-based evaluations. The project also includes a secondment at VANDAL (Politecnico di Torino), enabling further collaboration with leading VPR experts and fostering deeper insights into the proposed research tasks. Overall, this work aligns with the EU Work Programme on trustworthy, safe, and resilient AI and robotics systems. By enhancing both the efficiency and robustness of localization technologies, PAVPR will strengthen Europe’s leadership in trustworthy AI and deliver significant impact to the scientific community and industrial stakeholders alike.