With the increasing number of image-based learning datasets of places (e.g. Google Landmarks) and the enthusiasm for the automation of mobile systems (autonomous navigation, drones, etc.), the research on image-based localization, and its multiple variations (landmark retrieval, place recognition, etc.), has been growing rapidly, with different objectives in terms of location precision and of source of information exploited. Most of the approaches rely on 2D images, for which visual feature extraction has been a well-known process for decades.
As point clouds become increasingly used in applications thanks to low cost LiDAR sensors (Ouster, Velodyne to name but two), the availability of 3D photogrammetric sites or the development of public (mapping agencies) or private (HERE) digitization surveys, point cloud based retrieval becomes more and more studied. Currently, one of the most frequent use case is the loop closure step during point cloud based structure from motion, but it is clear that the advent of such data begins to make popular the general problem of place recognition from point clouds, especially because these geometric data may provide representations more robust to scene variations than images. Since PointNetVLAD, several articles have been published in top ranked conferences on this topic. The increasing popularity of the problem can be addressed from various research angles in 3D and 2D, including: model efficiency, indexing and retrieval, multi-modality 2D/3D, semantic segmentation, registration, reconstruction, scaling, etc.
On the large scale aspect, being able to recognize a place at a country scale (or a city) is an important achievement for practical applications. Actually, the amount of data processed by the state-of-theart approaches is several orders of magnitude below image processing applications. Indeed, those methods are focused on the scale of an acquisition (city scale at most), while in parallel, recent other approaches are specially designed to handle very large point cloud on big data frameworks, such as surface reconstruction algorithms. The advent of very large scale data set (e.g. full country LiDAR coverage) with multi modality will draw even more attention to this issue.