A Survey on the Architecture of Visual Mamba

Mamba, as a novel architecture based on State Space Models (SSMs), has introduced a paradigm shift in computer vision by addressing the limitations of traditional Convolutional Neural Networks (CNNs) and Transformers. Compared to CNNs’ inadequacy in global context modeling and Transformers’ scalability challenges due to quadratic computational complexity, Mamba achieves efficient long-range dependency modeling through its selective scanning mechanism and hardware-aware linear-time algorithms. This architecture demonstrates performance comparable to existing models in tasks such as image classification, object detection, and segmentation while significantly reducing computational costs.In recent years, numerous studies have explored extending Mamba from natural language processing to vision domains, investigating its potential across general vision tasks, multimodal applications, medical image analysis, and remote sensing image interpretation. To comprehensively understand Mamba’s applications in vision, we conduct a systematic survey. This survey not only focuses on Mamba’s architectural innovations as a backbone in vision tasks but also analyzes its structural adaptations across diverse vertical domains.