Heterogeneous Radar Multi-session PR
For multi-session evaluation, spinning radar data from sequence 01 were used as the database, and 4D radar data from sequences 02 and 03 were used as queries. Quantitative results are presented in Table II, with qualitative analysis shown in Fig. 6 and recall@N curves depicted in Fig. 7. We have demonstrated the necessity of heterogeneous radar PR models beyond homogeneous radar systems by showing that existing SOTA methods fail to address the inherent challenges of heterogeneous radar.
Homogeneous Radar Multi-session PR
SHeRLoc was compared with Autoplace, MinkLoc3Dv2, and TransLoc4D for 4D radar PR, and with RaPlace, RadVLAD, and FFT-RadVLAD for spinning radar PR. The results are reported in Table II and Fig. 8, confirming SHeRLoc’s superior performance despite not being designed for homogeneous radar. For 4D radar PR, it demonstrates top performance across nearly all sequences. For spinning radar PR, SHeRLoc achieves an AR@1 of 0.993, surpassing all SOTA methods. Furthermore, PR curves indicate that SHeRLoc achieves high precision and recall for both radar types under diverse conditions, demonstrating robust performance for both homogeneous and heterogeneous radar.