EPSA-VPR: A lightweight visual place recognition method with an Efficient Patch Saliency-weighted Aggregator

Visual Place Recognition (VPR) is important in autonomous driving, as it enables vehicles to identify their positions using a pre-built database. In this domain, prior research highlights the advantages of recognizing and emphasizing high-saliency local features in descriptor aggregation for performance improvement. Following this path, we introduce EPSA-VPR, a lightweight VPR method incorporating a proposed Efficient Patch Saliency-weighted Aggregator (EPSA), additionally addressing the computational efficiency demands of large-scale scenarios. With almost negligible computational requirements, EPSA efficiently calculates and integrates the local saliency into the global descriptor. To quantitatively evaluate the effectiveness, EPSA-VPR is validated across various VPR benchmarks. The comprehensive evaluations confirm that our method outperforms existing advanced VPR technologies and achieves competitive performance. Notably, EPSA-VPR also derives the second-best performance among two-stage VPR methods, without the need for re-ranking computations. Moreover, the effectiveness of our model is sustainable even with considerable dimension reduction. Visualization analysis reveals the interpretability of EPSA-VPR that after training, the backbone network learns to attach more attention on the task-related elements, which makes the final descriptor more discriminative.