RealX3D provides multi-view RGB images in JPEG and Sony RAW formats, COLMAP sparse reconstructions, and high-precision 3D ground-truth geometry. The benchmark includes a diverse set of scenes and challenging degradation types like low light and smoke for visual restoration and reconstruction tasks.
Use Cases
- Train multi-view 3D reconstruction models using COLMAP sparse reconstructions and high-precision ground-truth point clouds.
- Develop image restoration algorithms for low light or smoke conditions using the provided multi-view RGB images.
- Benchmark depth estimation models against the dataset's rendered depth maps.
- Compare reconstruction quality between processed JPEG and Sony RAW image sources.
Strengths
- Includes high-precision 3D ground-truth geometry in multiple formats: point clouds, meshes, and rendered depth maps.
- Provides multi-view RGB images in both processed JPEG and Sony RAW formats for source comparison.
- Covers a diverse set of scenes and real-world degradation types like low light and smoke.
Limitations
- Unknown sample size, scene count, and data volume limits statistical assessment.
- Unknown temporal and geographic coverage may restrict generalizability of trained models.
Provenance
- Source
- Hugging Face dataset by author ToferFish.
- Collection Method
- Real-world benchmark dataset for multi-view 3D reconstruction under challenging capture conditions.
- Time Range
- null
- Freshness
- Last updated March 2026.
- Geography
- null