The Renaissance of Classic Feature Aggregations for Visual Place Recognition in the Era of Foundation Models

Bingxi Liu, Pengju Zhang, Li He, Calvin Chen, Shiyi Guo, Yihong Wu, Jinqiang Cui, Hong Zhang

Keywords: Visual Place Recognition, Feature Representation, Supervised Learning

TL;DR: In contrast to recent works in ECCV'24, CVPR’24, and ICLR’24, we achieve state-of-the-art results in visual place recognition by simply modifying two methods from ten years ago.

Abstract: Visual Place Recognition (VPR) addresses the retrieval problem in large-scale geographic image databases through feature representations. Recent approaches have leveraged visual foundation models and have proposed novel feature aggregations. However, these methods have failed to grasp the core concepts of foundational models, such as leveraging extensive training sets, and have also neglected the potential of classical feature aggregations, such as GeM and NetVLAD, for low-dimensional representations. Building on these insights, we revive classic aggregation methods and create more fundamental VPR models, abbreviated SuperPlace. First, we introduce a supervised label alignment method that combines grid partitioning and local feature matching. This allows models to be trained on diverse VPR datasets within a unified framework, similar to the design principles of foundation models. Second, we introduce G$^2$M, a compact feature aggregation with two GeMs, in which one GeM learns the principal components of feature maps along the channel direction and calibrates the other GeM's output. Third, we propose the secondary fine-tuning (FT$^2$) strategy for NetVLAD-Linear (NVL). NetVLAD first learns feature vectors in a high-dimensional space and then compresses them into a low-dimensional space using a single linear layer. G$^2$M excels in large-scale applications requiring rapid response and low latency, while NVL-FT$^2$ is optimized for scenarios demanding high precision across a broad range of conditions. Extensive experiments (12 test sets, 14 previous methods, and 11 tables) highlight our contributions and demonstrate the superiority of SuperPlace. Specifically, SuperPlace-G$^2$M achieves state-of-the-art results with only one-tenth of the feature dimensions compared to recent methods. Moreover, SuperPlace-NVL-FT$^2$ holds the top rank on the MSLS challenge leaderboard. We have submitted a ranking screenshot, the source code, and the original experimental records in the supplementary materials.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2153