Reproduce “CVGL: Causal Learning and Geometric Topology” 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
Songsong Ouyang, Yingying Zhu
The CLGT framework integrates causal learning and geometric topology fusion for cross-view geo-localization, enhancing robustness against environmental variations and extreme viewpoint differences. It achieved a Recall@1 of 98.85% on CVUSA and improved Recall@1 by 6.62 percentage points on CVUSA-C-ALL, outperforming baselines on standard and corruption-augmented datasets.
Key insights
- Applying causal intervention in the frequency domain can effectively guide CVGL models to focus on stable, discriminative structural features by mitigating the influence of non-causal visual confounders.
- Robustly fusing BEV road topology information at the feature level provides critical geometric awareness, significantly improving alignment and consistency between disparate street and aerial views.
- An adaptive pooling mechanism, dynamically combining multiple pooling strategies, can enhance the final feature representation by better capturing and emphasizing semantically rich regions across varied scenes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.