SatUnreal is released under the SatUnreal Dataset License (Non-Commercial, No-Derivatives). Access is granted for educational or research use within a non-commercial institution only. For commercial licensing, contact khanai@telepix.co.kr.

SatUnreal: A High-Precision Synthetic Dataset for Satellite Stereo Matching via Unreal Engine

Paper | PDF | arXiv | Supplemental | GitHub

Han-Gyeol Kim* | JaeWan Park* | Junmin Park | Darongsae Kwon

TelePIX Inc., South Korea — _{* Equal contribution}

Presented at CVPR Workshop EarthVision 2026.

Dataset Summary

SatUnreal is a synthetic satellite stereo dataset with high-precision ground-truth disparity, rendered in Unreal Engine 5. It provides 10,000 disparity sets across 8 photorealistic 3D environments and 25 pivot points, generated at an altitude of 1 km with a ground sample distance of approximately 0.3 m and baselines ranging from 50 m to 200 m.

Unlike real-world benchmarks such as US3D and WHU-Stereo, SatUnreal is free from spatio-temporal mismatch (environmental change and shadow displacement between multi-view acquisitions) and provides unambiguous ground truth in occluded regions.

Disparity labels are produced by a reprojection-based geometric pipeline rather than approximate stereo formulas, with binary occlusion masks validated by a two-step line trace between the stereo cameras.

Dataset Structure

Total size: 63.29 GB

Disparity_All/
└── <Region>/
    ├── Disparity_Normalized/   # min-max normalized to [0, 255], for visual inspection
    ├── Disparity_raw/O(n)/     # raw float disparity (horizontal pixel displacement)
    ├── Color/(O(n)_L, O(n)_R)/ # rectified stereo RGB pairs
    └── Json/O(n)/              # per-sample camera and capture metadata

O(n) (Origin Target Point) is the ground-level scene center observed by the simulated sensor, defining the look-at pivot for each stereo capture configuration.

Region	Type	Pivot Points	Total Pairs	Size
California	Urban	8	3,200	22.5 GB
Toronto	Urban	10	4,000	21.3 GB
Venice	Urban	2	800	3.97 GB
Forest	Natural	1	400	2.58 GB
Canyon	Natural	1	400	2.89 GB
Coast	Natural	1	400	2.95 GB
Desert	Natural	1	400	2.5 GB
Field	Natural	1	400	4.6 GB
Total		25	10,000	63.29 GB

All regions use a baseline of 50–200 m (20 steps) and azimuth α ∈ [-90°, 90°] (20 random samples).

Download

Each region is packaged as a single zip under data/ (data/California.zip, data/Toronto.zip, ...); each archive extracts to <Region>/ with the layout above. Access is gated: sign in, accept the license on this page, then download from the Files and versions tab or with the CLI:

hf auth login
hf download telepix-ai/SatUnreal --repo-type dataset --local-dir ./SatUnreal                          # everything
hf download telepix-ai/SatUnreal data/Forest.zip --repo-type dataset --local-dir ./SatUnreal          # one region

Camera Metadata

Each stereo sample ships with a JSON file containing baseline_m, convergence_deg, gsd_m, target_pos_m, image resolution, and full intrinsics/extrinsics for both cameras (world_location_m, world_rotation_deg, fx_px, fy_px, cx_px, cy_px, sensor_width_mm, focal_length_mm). See the GitHub README for the full schema.

Usage Notes

⚠️ Important: The non-overlapping black regions (out-of-bounds areas) between the left and right stereo views must be cropped out before training. Using the raw images without removing these invalid regions may degrade model performance.

Licensing

This dataset is released under the SatUnreal Dataset License (Non-Commercial, No-Derivatives) — see the LICENSE file in this repository.

Use is permitted for non-commercial purposes only, meaning educational or research use solely within a non-commercial institution, in unmodified form. For any other use, including commercial licensing, contact khanai@telepix.co.kr.

Citation

@InProceedings{Kim_2026_CVPR,
    author    = {Kim, Han-Gyeol and Park, JaeWan and Park, Junmin and Kwon, Darongsae},
    title     = {SatUnreal: A High-Precision Synthetic Dataset for Satellite Stereo Matching via Unreal Engine},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2026},
    pages     = {7990-7999}
}

Acknowledgement

Evaluated with RAFT-Stereo, Selective-IGEV, and DLNR. Related datasets: US3D, WHU-Stereo.

Downloads last month: 36

Total file size:

62.6 GB

Paper for telepix-ai/SatUnreal

SatUnreal: A High-Precision Synthetic Dataset for Satellite Stereo Matching via Unreal Engine

Paper • 2609.27442 • Published 23 days ago