Reflect3r: Synthetic 3D Stereo Reconstruction Scenes with Mirror Reflections

A grouped dataset from FiftyOne containing the synthetic evaluation benchmark for Reflect3r, a 3DV 2026 paper on single-view 3D reconstruction. It includes 16 synthetic Blender interior scenes, each rendered from a real camera and a derived virtual mirror camera, accompanied by ground-truth point clouds. The dataset was created by Voxel51 and last updated on March 17, 2026.

Use Cases

• Benchmarking 3D reconstruction algorithms based on the synthetic scenes and ground-truth point clouds.
• Training models for mirror-aware scene understanding based on the paired real and virtual mirror camera views.
• Developing novel-view synthesis techniques based on the geometrically derived virtual camera perspectives.
• Evaluating depth estimation performance in indoor environments with reflective surfaces.

Strengths

• Contains 16 distinct synthetic interior scenes, providing a controlled evaluation environment.
• Includes geometrically derived virtual mirror camera views alongside real camera views for each scene.
• Provides ground-truth 3D geometry in the form of point clouds for validation.

Limitations

• Dataset is synthetic, which may limit generalizability to real-world, noisy camera data.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file formats are unknown, which may limit suitability assessment.

Provenance

Source

Voxel51, based on the synthetic benchmark from the Reflect3r research paper (3DV 2026).

Collection Method

Synthetic scenes generated and rendered using Blender.

Time Range

The benchmark is associated with a paper accepted for 2026.

Freshness

Last updated 2026-03-17 18:24:28

Geography

Synthetic data; no real-world geography.