Robust Visual Place Recognition

Why it matters

Place recognition is the backbone of loop closure in SLAM, of re-localisation after a robot gets lost, and of large-scale visual search. Patch-NetVLAD (CVPR 2021) showed how multi-scale local and global descriptors can be fused. Our IJCAI survey and IEEE Robotics and Automation Magazine tutorial helped define the field.

With an Amazon Research Award on Complementarity-Aware Multi-Process Fusion for Long Term Localization, we developed a hierarchical fusion framework that characterises where each localisation technique works well and fails, so that cheaper and more lightweight sensing and compute can still deliver long-term, privacy-preserving localisation for consumer robots. More recent work stacks descriptors into hyperdimensional place signatures, learns discriminative projections for multi-reference maps, and extends place recognition to LiDAR, aerial imagery in Antarctica, and comparisons with natural systems. How a system can tell when to trust its own answer is the focus of our introspection work.