Deep Multi View Stereo

Deep Multi-View Stereo (MVS) aims to reconstruct high-fidelity 3D models from multiple 2D images, focusing on improving accuracy and efficiency, particularly in challenging scenarios with sparse views or domain shifts. Current research emphasizes novel loss functions, transformer-based architectures for efficient high-resolution depth estimation, and iterative refinement algorithms that leverage geometric constraints like epipolar lines and surface normals, often incorporating confidence maps for improved robustness. These advancements are driving progress in areas like autonomous navigation, augmented reality, and cultural heritage preservation by enabling more accurate and efficient 3D scene capture and modeling.