WMGStereo: 150,000 Synthetic Stereo Pairs for Zero-Shot Matching

WMGStereo-150k provides 150,000 synthetic stereo image pairs and disparity maps generated by the Princeton Vision & Learning Lab. Released in 2025, the collection uses procedural generation to create indoor, nature, and dense "flying" scenes for depth estimation tasks.

Use Cases

• Training stereo matching models using the 150,000 synthetic image pairs
• Evaluating zero-shot transfer capabilities using the diverse indoor and nature scene categories
• Supervised depth estimation training using the provided ground-truth disparity maps

Strengths

• 150,000 high-resolution synthetic training samples
• Procedural generation covering three distinct scene types: indoor, nature, and flying objects
• BSD 3-Clause license allows for flexible academic and commercial use

Limitations

• Synthetic-to-real domain gap may affect performance on physical camera data
• Procedural nature may lack the specific geometric edge cases found in manual 3D modeling

Provenance

Source

Princeton Vision & Learning Lab (Arxiv 2504.16930)

Collection Method

synthetic

Freshness

Last updated March 2026