ANU-RL: A New Perspective on Unsupervised Representation Learning for Visual Place Recognition

Abstract: Representation Learning (RL) is fundamental for image matching, retrieval, classification, and other applications, enabling task-specific feature learning. RL algorithms aim to learn compact embeddings that preserve the neighbourhood structure of the input data. A general approach to this is contrastive learning, which pulls similar images (positives) closer together and pushes dissimilar images (negatives) farther apart in the embedding space. In Visual Place Recognition (VPR), positive images of a query share specific geographical and visual attributes with the query and can, form a cluster. In contrast, negative images differ from the query and may vary among themselves or be similar. % may vary or be similar among themselves. Most existing training objectives focus only on the relationships between query-positives and query-negatives. In this work, we hypothesize that, in addition to these relationships, other naturally available relationships, such as positives-to-negatives and intra-positives, can improve VPR performance by enhancing representation quality. The proposed framework, A New Perspective on Unsupervised Representation Learning (ANU-RL), when integrated with VPR aggregators like BoQ, SALAD, MixVPR, and NetVLAD, achieves state-of-the-art performance on most challenging VPR benchmarks, including Pittsburgh 30k, Tokyo 24/7, Nordland, MSLS (val), and many others. Moreover, all of this comes at no extra cost at inference time. Further, we generalize the proposed framework to a wider range of metric learning applications, specifically image retrieval.

Submission Type: Regular submission (no more than 12 pages of main content)

Previous TMLR Submission Url: https://openreview.net/forum?id=Ac4Z4ivA4N&referrer=%5BAuthor%20Console%5D(%2Fgroup%3Fid%3DTMLR%2FAuthors%23your-submissions)

Changes Since Last Submission: Changed the font as per the TMLR format

Assigned Action Editor: ~Yannis_Kalantidis2

Submission Number: 8585