Urban3D is a large-scale dataset of real-world, object-centric multiview videos and 3D reconstructions, captured to advance machine learning research in autonomous systems, smart infrastructure, and urban scene understanding.

The dataset contains over 10,000 multiview videos across urban categories, including traffic signage, dumpsters, construction equipment, e-scooters, road hazards, and more.

Each object in the dataset is derived from a single video sequence processed through a custom pipeline. This pipeline extracts high-resolution RGB frames, performs Structure-from-Motion (SfM) using COLMAP, and generates key assets for 3D reconstruction and neural rendering. Outputs include COLMAP-derived camera poses, a COLMAP database with feature matches, sparse 3D point clouds, and distortion-aware intrinsics and extrinsics. An instance-ready transforms.json is provided for NeRF pipelines, and the dataset supports Gaussian Splatting through optimized SIFT extraction and optional image undistortion. The original video file is retained for traceability and simulation. This structure ensures clean, calibrated geometry for NeRF training, Gaussian splat generation, and broader simulation workflows.