Reflect3R: Synthetic Data for Single-View 3D Reconstruction with Mirrors

Reflect3R is a synthetic dataset for training models to perform 3D stereo reconstruction from a single camera view, aided by mirror reflections. The data was created by researchers from the University of Oxford and is associated with a paper for the 3DV 2026 conference. It was last updated on the Hugging Face platform in March 2026.

Use Cases

• Training neural networks for single-view 3D reconstruction based on the described task.
• Benchmarking novel 3D stereo reconstruction algorithms against the synthetic data.
• Studying the role of mirror reflections as an aid for depth perception in monocular systems.

Strengths

• Data is associated with a peer-reviewed academic paper for 3DV 2026.
• Platform tags indicate the data includes both images and point clouds, suggesting a multimodal structure.

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 and file size are unknown, which may limit suitability assessment.

Provenance

Source

University of Oxford researchers (Jing Wu, Zirui Wang, Iro Laina, Victor Adrian Prisacariu).

Collection Method

Synthetic data generation, likely using Blender as indicated by platform tags.

Freshness

Last updated 2026-03-05 12:24:57; freshness should be verified.