Reproduce “Cross-view Geo-localization with Evolving Transformer” on OpenResearch

About this paper

Hongji Yang, Xiufan Lu, Yingying Zhu

The Evolving Geo-localization Transformer (EgoTR), an architecture developed for cross-view geo-localization, adapts the Transformer with learnable positional embeddings and a novel self-cross attention mechanism. It achieves state-of-the-art recall accuracy on CVACT and CVUSA datasets, while also producing more efficient image descriptors and demonstrating improved generalization across distinct datasets.

Key insights

• Transformer architectures can effectively model global dependencies and learn flexible position-aware representations, making them suitable for handling large viewpoint differences in cross-view geo-localization.
• Learnable positional embeddings enable the network to intrinsically capture relative geometric information, reducing the need for explicit geometric pre-alignment techniques like polar transform.
• Modeling correlations between non-consecutive layers via self-cross attention promotes the evolution and diversification of learned representations, leading to enhanced generalization ability and more stable network optimization.

Summary by alphaXiv.