Dataset Card for NL3D-Synth-Benchmark

Dataset Summary

NL3D-Synth-Benchmark is the synthetic evaluation split of NL3D, a synthetic-real dataset and benchmark for non-Lambertian 3D reconstruction. This split is designed to provide a compact, controlled, and reproducible benchmark for evaluating 3D perception methods on challenging non-Lambertian materials.

The benchmark contains 16 synthetic scenes rendered with physically based material models and full 3D background context. Each scene includes multi-view images, camera parameters, object masks, material information, and precise geometry annotations. The dataset focuses on four controlled non-Lambertian material types:

• conductor: metallic conductor materials;
• dielectric transmissive: transmissive dielectric materials;
• dielectric reflective: opaque reflective dielectric materials;
• dielectric subsurface: dielectric materials with subsurface scattering.

This synthetic benchmark is intended for controlled evaluation rather than large-scale training. For large-scale model pre-training or fine-tuning, please refer to the separate NL3D-Synth-Train split.

Dataset Links

NL3D is released as multiple dataset components:

Component	Description	Link
NL3D-Synth-Train	Large-scale synthetic training split with over 3,000 objects and over 8,000 scenes for pre-training and fine-tuning general-purpose 3D models.	NL3D-Synth-Train
NL3D-Synth-Benchmark	Compact synthetic benchmark with 16 controlled scenes for evaluating non-Lambertian 3D perception.	NL3D-Synth-Benchmar
NL3D-Real-Benchmark	Real-world benchmark containing more than ten captured scenes of transparent and metallic objects.	NL3D-Real-Benchmark
Project Page	Dataset documentation, benchmark protocol, examples, and updates.	[TODO: add URL]
Code / Evaluation Toolkit	Scripts for loading data, running evaluation, and reproducing benchmark results.	[TODO: add URL]

Supported Tasks

The data and annotations in NL3D-Synth-Benchmark can support multiple 3D perception tasks, including but not limited to:

• novel-view synthesis;
• geometry reconstruction;
• camera pose estimation;
• material-aware reconstruction;
• multi-view geometric analysis;
• evaluation of feed-forward 3D reconstruction models.

In the accompanying benchmark, we focus on three representative tasks:

1. Novel-view synthesis: evaluating whether methods can synthesize images from held-out camera viewpoints under complex view-dependent appearance.
2. Geometry reconstruction: evaluating the accuracy and completeness of reconstructed 3D geometry.
3. Camera pose estimation: evaluating the robustness of pose estimation methods on non-Lambertian scenes where feature matching and photometric consistency may be affected by reflection, transmission, refraction, or subsurface scattering.