Cross-view Transformer for enhanced multi-view 3D reconstruction

3D reconstruction from multiple 2D images provides rich interactive experiences in design, entertainment, and robotics. However, effectively fusing complementary information across viewpoints remains challenging. This paper proposes a novel cross-view Transformer-based approach for multi-view 3D reconstruction. Our method introduces a cross-view Transformer encoder that achieves effective interaction of information across views. We also develop a global-aware token fusion module to compress multi-view features and an agent attention-based decoder to reduce complexity while maintaining high reconstruction performance. Experiments on benchmark datasets demonstrate that our method achieves state-of-the-art results, outperforming existing techniques.