MVMamba: A Multi-View Mamba with Tree Scanning Strategy for Remote Sensing Semantic Segmentation Task

Le résumé fourni par la source

Semantic segmentation is a critical yet challenging task in remote sensing (RS). Transformer-based models dominate the field but often face high computational demands, requiring a trade-off between accuracy and efficiency. To address this challenge, we propose Multi-View Mamba (MVMamba), a novel model that balances performance and efficiency by combining tree topology scanning and multi-scale feature processing. MVMamba’s encoder extracts local features using convolutional neural networks (CNNs) of varying kernel sizes, while its decoder employs three branches: CNNs for local features, multi-scale tree-Mamba blocks for midlevel features, and attention mechanisms for global context. The tree-Mamba block organizes spatially related pixels for topology-aware scanning and adaptive input handling. Experiments on the ISPRS Vaihingen and LoveDA Urban datasets demonstrate that MVMamba outperforms state-of-the-art models, showcasing its efficiency and effectiveness in RS semantic segmentation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.