Expanded validation

Robustness against Random Rotation

• In real-world scenarios, random yaw angle rotations between the database and query are common, making rotation robustness crucial. Therefore, we conducted experiments with random yaw angle rotations on database scans within a 360 degree range. SHeRLoc demonstrates superior rotation robustness compared to other methods. As shown in Table III, comparing AR@1 with SOTA methods on Sports Complex, Library, and River Island, both SHeRLoc and SHeRLoc-S consistently achieve the best performance. This robust performance, despite angle variations, demonstrates that a limited number of multi-views is sufficient for achieving robust rotation invariance.

Zero-shot Performance on Unseen Datasets

• We evaluated SHeRLoc’s zero-shot generalization on unseen datasets, including MulRan and the Oxford Radar RobotCar. While HeRCULES employs the Navtech RAS6 spinning radar, MulRan uses the earlier Navtech CIR204-H, and Oxford Radar RobotCar uses the Navtech CTS350-X. These hardware differences induce substantial domain shifts in terms of noise characteristics, resolution, and detection range. For MulRan, we conducted single-session PR on the DCC 01 and KAIST 03 sequences. For Oxford Radar RobotCar, we performed multi-session PR using 2019-01-18 (#3) as the database and 2019-01-10 (#1) and 2019-01-16 (#2) as queries. As shown in Table IV, SHeRLoc achieves strong recall even under challenging zero-shot settings with domain shifts.

LiDAR to Radar

• We evaluated heterogeneous range sensor PR performance using FMCW LiDAR queries and a spinning radar database, comparing it with Radar Scan Context, Autoplace, and Radar-to-LiDAR. We utilized radial velocity in a manner analogous to 4D radar, and replaced RCS with reflectivity ρ. Our preprocessing aligns the distributions of FMCW LiDAR and spinning radar, and Table V confirms that SHeRLoc remains robust across heterogeneous range sensors.