Multi-View Depth Estimation

Multi-view depth estimation uses images from different viewpoints to infer scene depth, resolve scale ambiguity, and build consistent 3D geometry for applications such as mapping, robotics, and autonomous driving. Classical multi-view stereo methods use camera calibration, feature matching, plane-sweep cost volumes, and geometric optimization, while newer transformers use camera-aware features and cross-view attention to combine evidence across images. Many systems also incorporate monocular depth foundation-model priors, sparse LiDAR anchors, temporal video cues, or self-supervised consistency to handle sparse views, occlusions, and changing scenes. A central challenge is balancing accurate, globally consistent geometry with efficiency and generalization across camera types, domains, and dynamic environments.