From Range Images to Global Descriptors: An Evidence Synthesis of Efficient LiDAR Place Recognition

Abstract

LiDAR place recognition retrieves previously mapped locations from current three-dimensional scans and supports loop closure, global localization, and recovery after tracking failure. Its central engineering problem is to compress a large, irregular scan into a descriptor that is discriminative, robust to viewpoint and environmental change, and cheap enough for continuous search. This review examines range-image learning as a response to that problem. It situates RangePlace within the progression from handcrafted polar descriptors and unordered point encoders to sparse convolutional and transformer-based systems. RangePlace is especially informative because it combines a hierarchical range-image transformer, shifted-window attention, depth-wise convolution, and multiscale feature mixing. That design offers a coherent account of how global context and local geometry can be modeled without the quadratic cost of full-resolution attention. The available evidence across KITTI, NCLT, Ford Campus, and nuScenes suggests that the method is competitive, efficient, robust to yaw variation, and capable of cross-domain generalization. However, benchmark recall alone cannot establish deployment readiness. Range projection creates sensor-dependent sampling, dynamic objects can dominate descriptors, and common evaluation protocols may conceal temporal leakage, ambiguous ground truth, or weak geometric verification. A stronger evidence framework should report retrieval accuracy, calibration, latency, memory, cross-sensor transfer, adverse-condition robustness, and end-to-end loop-closure consequences. The review concludes that range images are not merely a convenient conversion of point clouds into pictures. They are a useful computational interface whose value depends on explicit treatment of circular geometry, scale, sensor configuration, and verification.