Structured3D

Why we included this project

Structured3D is a synthetic dataset of 3,500 house designs, created by professional designers and rendered into photo-realistic images, each carrying dense ground-truth structure annotations such as depth, surface normals, semantic maps, and room layouts under varied lighting and furniture. It is aimed at people working on indoor scene understanding, room-layout estimation, or monocular 3D reconstruction, who need consistent ground truth to train and evaluate models that recover structured 3D from 2D images. Because the annotations come from known 3D geometry rather than hand labeling, they are consistent and complete. The repo also ships small Python viewers for inspecting wireframes, planes, floorplans, and bounding boxes, so you can sanity-check predictions against the ground truth. One practical caveat: the data is gated behind a terms-of-use agreement, so plan for that when budgeting a project.