<p>3D reconstruction from multiple images, known as Multi-View Stereo (MVS), remains a challenging problem in computer vision, particularly when balancing reconstruction accuracy with computational efficiency. While most existing methods adopt CNN-based architectures that primarily focus on extracting local contextual features, they often fail to incorporate global context effectively resulting in limited performance in complex scenes. To address this limitation, we propose a novel MVS framework that integrates global contextual information into both the 2D and 3D stages of the reconstruction pipeline. Specifically, we introduce an Atrous Spatial Pyramid Pooling (ASPP) block in the 2D feature extraction stage and a 3D-ASPP block after the cost volume regression stage. These modules are designed to capture multi-scale contextual features while maintaining a low memory footprint and minimal computational overhead. Extensive experiments on two benchmark datasets DTU and Tanks and Temples demonstrate that our method consistently outperforms existing CNN based and Transformer based models in terms of reconstruction quality and training convergence speed. Our approach achieves a favorable trade-off between accuracy and efficiency, making it a practical solution for real-world 3D reconstruction applications.</p>