EffoVPR: Effective Foundation Model Utilization for Visual Place Recognition

TL;DR

EffoVPR tackles Visual Place Recognition by exploiting pre-trained foundation models to produce compact global descriptors and a lightweight, zero-shot re-ranking mechanism. It uses the ViT self-attention outputs from intermediate layers as local keypoints/descriptors and relies on mutual nearest-neighbor matching to refine top candidates, with fixed thresholds T1 and T2 and no external pooling modules. The approach demonstrates state-of-the-art results on multiple benchmarks, achieving strong robustness to occlusion, day-night, and seasonal changes, and shows that fine-tuning only a small tail of the backbone further boosts performance. The work offers a practical, memory-efficient solution for large-scale VPR, combining zero-shot capability with optional fine-tuning to reach top performance while maintaining feature compactness.

Abstract

The task of Visual Place Recognition (VPR) is to predict the location of a query image from a database of geo-tagged images. Recent studies in VPR have highlighted the significant advantage of employing pre-trained foundation models like DINOv2 for the VPR task. However, these models are often deemed inadequate for VPR without further fine-tuning on VPR-specific data. In this paper, we present an effective approach to harness the potential of a foundation model for VPR. We show that features extracted from self-attention layers can act as a powerful re-ranker for VPR, even in a zero-shot setting. Our method not only outperforms previous zero-shot approaches but also introduces results competitive with several supervised methods. We then show that a single-stage approach utilizing internal ViT layers for pooling can produce global features that achieve state-of-the-art performance, with impressive feature compactness down to 128D. Moreover, integrating our local foundation features for re-ranking further widens this performance gap. Our method also demonstrates exceptional robustness and generalization, setting new state-of-the-art performance, while handling challenging conditions such as occlusion, day-night transitions, and seasonal variations.

Figures (10)

• Figure 1: EffoVPR showcase. Top row: we show a challenging query image (a) and EffoVPR's top-1 candidate (b), retrieved from a gallery of 2.8M geo-tagged images of SF-Occ. EffoVPR demonstrates high capability in handling transitional objects and large obstructions. Bottom row: We present Recall@1 performance of EffoVPR using global features against feature dimensionality. While the current leading methods achieve their performance using large features, EffoVPR demonstrates top performance even with an extremely compact feature size.
• Figure 1: Ablation on the number of trainable layers
• Figure 2: An overview of EffoVPR. Left: During inference, we identify the nearest neighbors of the query by using the $\text{[CLS]}$ token as the global representation for each image ($v_g$). For the second re-ranking stage, we extract (dashed-line) intermediate features, and utilize $S$, the partial attention map, for keypoint selection (with a predefined threshold $T_1$), while employing the Value facet $V$ as the corresponding keypoints descriptors. Right: Lastly, we re-rank the top-K candidates from the first stage based on the count of strongly connected mutual nearest neighbors (MNN) with a score exceeding a predefined threshold (denoted as $T_2$).
• Figure 2: Recall@1 performance of EffoVPR-G global feature versus feature dimensionality for more datasets.
• Figure 3: EffoVPR zero-shot. (a) Comparison of EffoVPR-ZS with other VPR trained methods. Our zero-shot approach shows comparable results. (b) Zero-shot success despite existing dynamic and irrelevant objects and strong visual change. Matching keypoints are indicated by colored lines. Although the pre-trained DINOv2 initially has its strongest attention on the distracting temporal advertisement, EffoVPR effectively identifies correct keypoints for successful matching.