UnrealMVS: Synthetic Omnidirectional Depth Dataset for Fisheye Camera Rigs

UnrealMVS is a large-scale synthetic omnidirectional depth dataset rendered in Unreal Engine. It is designed for training and evaluating multi-view stereo networks, such as OmniMVS, on fisheye camera rigs. The dataset was created by flw-tu-dortmund and was last updated on 2026-05-22.

Use Cases

• Train multi-view stereo networks based on synthetic fisheye image pairs.
• Benchmark depth estimation algorithms based on provided ground-truth depth maps.
• Develop and evaluate models for 3D reconstruction from omnidirectional imagery based on the described camera rig setup.

Strengths

• Provides ground-truth depth maps in meters, which are essential for supervised learning.
• Uses a consistent capture setup with a 4-fisheye-camera rig per scene.
• Images are stored in a high dynamic range format (EXR), likely preserving more detail.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

flw-tu-dortmund

Collection Method

Rendered synthetically in Unreal Engine.

Freshness

Last updated 2026-05-22 19:01:45; freshness should be verified.