What is ivgt?

IVGT stands for Implicit Visual Geometry Transformer, a research project in computer vision that reconstructs three dimensional scenes from multiple photographs taken without any prior knowledge of where the camera was positioned. Most competing approaches work by assigning a 3D point to every pixel in every photo, called explicit geometry, which means the same surface gets predicted multiple times from different angles, leading to inconsistencies and gaps. IVGT takes a different approach: instead of per pixel predictions, it learns a continuous three dimensional field, called a signed distance function, that can be queried at any point in space to determine whether that point is on, inside, or outside a surface. Because the field is continuous rather than pixel by pixel, IVGT can extract complete, coherent 3D meshes, the polygon surfaces used in 3D graphics, in a single pass through the model, with no slow per scene optimization required. The same underlying field also produces depth maps showing how far each surface is from the camera, surface normal maps showing which direction each surface faces, and rendered images from new viewpoints the camera never captured. The README reports benchmark comparisons against other recent reconstruction models on tasks including mesh reconstruction, point cloud reconstruction, novel view synthesis, depth estimation, surface normal estimation, and camera pose estimation, with IVGT performing competitively or best on several of them. The model is trained across multiple datasets using only 2D image supervision and geometric constraints, allowing it to generalize to new scenes without further tuning. According to the README, the code has not yet been released publicly, only the paper's findings and results are shown. This project is aimed at computer vision researchers working on 3D reconstruction, novel view synthesis, or scene understanding.