FoL++ Visual Place Recognition Training Pipeline
Figure 2: Overview of the FoL++ training pipeline. The DINOv2 backbone extracts patch tokens, which are aggregated via Sinkhorn optimal assignment to form a robust global descriptor D g D_{g} . In parallel, a lightweight decoder recovers high-resolution local features D l D_{l} . Meanwhile, the Reliability Estimation Branch generates a spatial reliability map 𝐑 \mathbf{R} , which is fused with the global assignment mask 𝐌 a \mathbf{M}_{a} to yield the final discriminative mask 𝐌 \mathbf{M} . Local features are optimized via weakly supervised pseudo-correspondences. The entire framework is trai
Paper context
Paper title: Region Matters: Efficient and Reliable Region-Aware Visual Place Recognition Abstract: Visual Place Recognition (VPR) determines a query image's geographic location by matching it against geotagged databases. However, existing methods struggle with perceptual aliasing caused by irrelevant regions and inefficient re-ranking due to rigid candidate scheduling. To address these issues, we introduce FoL++, a method combining robust discriminative region modeling with adaptive re-ranking. Specifically, we propose a Reliability Estimation Branch to generate spatial reliability maps that explicitly model occlusion resistance. This representation is further optimized by two spatial alignment losses (SAL and SCEL) to effectively align features and highlight salient regions. For weakly supervised learning without manual annotations, a pseudo-correspondence strategy generates dense local feature supervision directly from aggregation clusters. Our Adaptive Candidate Scheduler dynamically resizes candidate pools based on global similarity. By weighting local matches by reliability and adaptively fusing global and local evidence, FoL++ surpasses traditional independent matching systems. Extensive experiments across seven benchmarks demonstrate that FoL++ achieves state-of-the-art performance with a lightweight memory footprint, improving inference speed by 40% over FoL. Code and models will be released (and merged with FoL) at https://github.com/chenshunpeng/FoL. Passages referencing this figure: eriments across seven benchmarks demonstrate that FoL++ achieves state-of-the-art performance with a lightweight memory footprint, improving inference speed by 40% over FoL. Code and models will be released (and merged with FoL) at https://github.com/chenshunpeng/FoL . keywords: Visual Place Recognition, Discriminative Region Learning, Weakly Supervised Learning, Foundation models. 1 Introduction Figure 1: Comprehensive evaluation of the proposed FoL++ framework. (a) Visual comparison between standard matching (top) and our discriminative region-guided matching (bottom), showing the effective suppression of mismatches in irrelevant backgrounds. (b) Trade-off analysis demonstrating that FoL++ achieves optimal performance across all metrics. All methods are evaluated on an NVIDIA A100 GPU at nd SelaVPR [ selavpr ] , suffer from inefficiencies. Although they utilize multi-layer dense features or upsampled descriptors to ensure broad spatial coverage, they typically perform exhaustive matching across the entire representation without explicitly filtering out background noise. This lack of discriminative region selection inevitably wastes computational resources on irrelevant areas (see Fig. 1 (a) and (b)). Furthermore, existing frameworks treat global similarity assessment and local verification as disconnected components [ patchvlad , r2former , effovpr ] , leading to suboptimal candidate ranking. Based on this analysis, we identify three core challenges inherent in current approaches: 1. Inadequate Spatial Reliability Modeling: Traditional methods often lack explicit mechanism aditional methods often lack explicit mechanisms to identify spatially stable regions against extreme viewpoint changes and severe occlusions. Although conceptually distinct, these challenges freque
