In the world of multi-view stereo reconstruction, obtaining camera parameters can be tedious but crucial. However, DUSt3R presents a groundbreaking approach that eliminates the need for prior camera calibration or viewpoint poses. By treating pairwise reconstruction as a regression of pointmaps, DUSt3R breaks free from traditional constraints and seamlessly combines monocular and binocular reconstruction. With a network architecture based on Transformer encoders and decoders, DUSt3R produces 3D models, depth information, pixel matches, and camera poses effortlessly. This all-in-one method sets new standards in 3D vision tasks, making complex geometric processes easier to achieve.
https://dust3r.europe.naverlabs.com/