HypeVPR: Exploring Hyperbolic Space for Perspective to Equirectangular Visual Place Recognition

TL;DR

This work introduces HypeVPR, a hierarchical embedding framework in hyperbolic space specifically designed to address the challenges of P2E matching, and proposes a hierarchical feature aggregation mechanism that organizes local-to-global feature representations within hyperbolic space.

Abstract

Visual environments are inherently hierarchical, as a panoramic view naturally encompasses and organizes multiple perspective views within its field. Capturing this hierarchy is crucial for effective perspective-to-equirectangular (P2E) visual place recognition. In this work, we introduce HypeVPR, a hierarchical embedding framework in hyperbolic space specifically designed to address the challenges of P2E matching. HypeVPR leverages the intrinsic ability of hyperbolic space to represent hierarchical structures, allowing panoramic descriptors to encode both broad contextual information and fine-grained local details. To this end, we propose a hierarchical feature aggregation mechanism that organizes local-to-global feature representations within hyperbolic space. Furthermore, HypeVPR's hierarchical organization naturally enables flexible control over the accuracy-efficiency trade-off without additional training, while maintaining robust matching across different image types. This approach enables HypeVPR to achieve competitive performance while significantly accelerating retrieval and reducing database storage requirements. Project page: https://suhan-woo.github.io/HypeVPR/

Figures (4)

• Figure 1: The perspective to equirectangular visual place recognition (P2E VPR) problem requires matching a distinct query image with a panoramic database image. The yellow-highlighted region (4th from the left) corresponds to the matching area, while the rest is dominated by redundant FoVs, making it difficult to generate a single representative descriptor. (a) In Euclidean space, modeling the global descriptor that faithfully preserves relational distances among multiple fields of view (FoVs) is challenging. (b) Even with hierarchical modeling, Euclidean representations suffer from geometric distortion that hinders accurate hierarchy preservation. (c) Our hierarchical modeling in hyperbolic space naturally embeds multi-level relationships with minimal distortion, producing a compact representative descriptor that effectively integrates all FoVs within the equirectangular image. (d) Comparison of speed, recall, and storage efficiency against state-of-the-art methods, showing the favorable trade-offs of our approach. (See Tab. \ref{['tab:sf_xl']} for details.)
• Figure 2: (a) Structure of our network and our training scheme. (b) Illustration of the Hierarchical Aggregation Module (HAM).
• Figure 3: Qualitative results. Correct predictions are outlined in green, while incorrect predictions are outlined in red. In our method, the window with the highest score within the positive pair is highlighted in yellow. Best viewed when zoomed in.
• Figure 4: Visualization of 1,000 hierarchical descriptors of Pitts-250k P2E test set on the Poincaré ball manifold.