FFS Stereo4D: 1.3 Million FoundationStereo Disparity Maps

NVIDIA's FFS Stereo4D contains approximately 1.3 million disparity maps generated from the Stereo4D dataset using the FoundationStereo model. These 784×784 PNG images provide per-pixel disparity values for stereo matching and depth estimation tasks. The collection is organized into 26 zip files, each containing 50,000 frames with associated metadata.

Use Cases

• Training depth estimation models using decoded disparity values from the RGB channels
• Evaluating stereo matching algorithms against FoundationStereo-generated labels
• Temporal 3D scene analysis using the vid_id and frame_idx sequences

Strengths

• Scale of approximately 1.3 million disparity maps
• High-resolution 784×784 pixel density for detailed depth mapping
• Precision encoding using 3-channel uint8 format to represent sub-pixel disparity

Limitations

• Labels are synthetic/model-generated by FoundationStereo rather than captured by physical sensors
• Restricted to non-commercial use under CC BY-NC 4.0 license

Provenance

Source

NVIDIA

Collection Method

synthetic

Freshness

Last updated March 2026