Anisotropic Synthetic Dataset (ASD)

Overview

The Anisotropic Synthetic Dataset (ASD) is a synthetic benchmark dataset originally introduced in Spec-Gaussian for evaluating neural rendering methods and inverse rendering algorithms on objects with pronounced specular reflections. This extended version was created to support the research presented in ShinyNeRF, providing comprehensive ground truth data for material and geometry recovery evaluation.

Dataset Description

ASD contains renderings of four distinct objects selected from the original dataset, each exhibiting distinctive specular characteristics where anisotropic material parameters remain roughly constant across the object's surface. While the original dataset provided high-quality RGB views, it did not include ground truth material properties and surface geometry. To enable quantitative benchmarking of material recovery, we re-rendered the scenes in Blender to generate complete ground truth data.

Key Features

• Varying geometric complexity: Diverse object shapes eliminate material ambiguity, allowing researchers to focus on algorithm robustness across different surface structures
• Homogeneous material parameters: Constant anisotropic properties per object enable controlled evaluation of material estimation
• Controlled illumination conditions: Direct point lighting provides predictable interactions without environment map complexity
• High-quality ground truth: Physically-based rendering provides accurate reference data for benchmarking material and geometry recovery

Dataset Contents

• RGB renderings: Rendered images (270×270 pixels) for each object under point light sources
• Material property maps: Ground truth maps including RGB, surface normals, tangent vectors, and depth (alpha)
• Blender scene files: Complete scene setups for reproducibility and custom rendering configurations