Reproduce “$R^{2}$Former: Unified $R$etrieval and $R$eranking Transformer for Place Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sijie Zhu, Linjie Yang, Chen Chen, Mubarak Shah, Xiaohui Shen, Heng Wang
R²Former combines transformer-based global retrieval with learned local reranking for visual place recognition, replacing conventional RANSAC verification with a correlation-aware transformer. Across MSLS, Pitts30k, Tokyo 24/7, Revisited San Francisco, and St Lucia, it improves localization accuracy while reducing reranking cost through attention-guided token selection and batched candidate processing.
Key insights
- Learned reranking benefits from combining local correlation with spatial coordinates and attention, rather than relying on geometric consistency alone; the local correlation signal is especially important in the ablations.
- Transformer patch tokens can support effective local matching despite global self-attention, and the reranking module can also operate with CNN-based features.
- The reranker is most useful after strong candidate retrieval: it substantially improves recall but cannot recover places missing from the initial top-k set.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.