LFVS-Mamba: State-space model for light field view synthesis

Abstract

Light Field View Synthesis (LFVS) methods using Convolutional Neural Networks (CNNs) and Vision Transformers (VTs) have been extensively studied: CNNs excel at learning local spatial features via hierarchical receptive fields but cannot capture long-range global dependencies, while VTs inherently model global context through self-attention at the cost of quadratic computation and memory complexity. To address these issues, we propose LFVS-Mamba, which integrates a State-Space Module (SSM) with a Selective Scanning Mechanism to efficiently capture long-range dependencies. LFVS-Mamba processes 2D slices of the 4D LF to fully exploit spatial context, complementary angular information, and depth cues. The LFVS-Mamba comprises three modules to progressively synthesize dense LFs: (i) Shallow Feature Extraction (SFE), (ii) Spatial-Angular Depth Feature Extraction (SADFE), and (iii) Angular Upsampling (AU). Experimental results on standard LF benchmarks demonstrate that LFVS-Mamba consistently outperforms existing methods.