OmniScene: Photorealistic Synthetic Scenes with Multi-Camera Depth Ground Truth

OmniScene is a photorealistic synthetic dataset rendered in NVIDIA Isaac Sim. It was created by henryzhou998 to train the X-Lens model and was last updated on July 11, 2026. Each scene is captured by a 6-camera rig consisting of 4 fisheye and 2 pinhole cameras, providing metric ground-truth depth.

Use Cases

• Train depth estimation models based on the provided metric ground-truth depth.
• Develop models for handling heterogeneous camera inputs based on the mix of fisheye and pinhole camera data.
• Benchmark multi-view 3D reconstruction algorithms using the photorealistic synthetic scenes.
• Research domain adaptation from synthetic to real-world imagery using the photorealistic renders.

Strengths

• Provides metric ground-truth depth for each scene, a precise form of annotation.
• Each sample contains data from a heterogeneous 6-camera rig (4 fisheye + 2 pinhole).
• Scenes are photorealistic, rendered using the NVIDIA Isaac Sim simulation platform.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data is synthetic, which may introduce a sim-to-real gap for model generalization.

Provenance

Source

henryzhou998 on Hugging Face

Collection Method

Rendered synthetically in NVIDIA Isaac Sim.

Freshness

Last updated 2026-07-11 15:06:12; freshness should be verified.