Download
Hypersim is a large-scale, photorealistic synthetic dataset and tooling suite for indoor scene understanding research. It provides richly annotated renderings—RGB, depth, surface normals, instance and semantic segmentations, and material/lighting metadata—produced from high-fidelity virtual environments. The dataset spans diverse furniture layouts, room types, and camera trajectories, enabling robust training for geometry, segmentation, and SLAM-adjacent tasks. Rendering pipelines and utilities allow researchers to reproduce sequences, generate novel views, or extract task-specific supervision. Because the data are perfectly labeled and controllable, Hypersim is well suited for pretraining and for studying domain transfer to real imagery. The repository acts as both a dataset index and a set of scripts for downloading, managing, and evaluating on standardized splits.
Features
- Photorealistic indoor scenes with multi-modal annotations
- Depth, normals, instance/semantic masks, and materials/lighting
- Reproducible rendering pipelines and camera trajectories
- Tools for dataset download, indexing, and standardized splits
- Strong pretraining source for geometry and segmentation tasks
- Controls for viewpoint and scene generation to study generalization
Project Samples
