Introduction
Accurately reconstructing objects with complex optical properties, such as metals and glass, remains a formidable challenge due to their unique specular and light-transmission characteristics. To facilitate the development of solutions to these challenges, we introduce the OpenMaterial dataset, comprising 1001 objects made of 295 distinct materials—including conductors, dielectrics, plastics, and their roughened variants— and captured under 723 diverse lighting conditions.
Overview
Our dataset contains 1001 object centered scenes. To eliminate scale bias, we standardize object sizes within a unit sphere. This lets us sample camera positions using a Fibonacci grid on the upper hemisphere, ensuring uniform distribution and non-overlapping coverage. We then splits these camera positions into distinct training (50) and testing (40) viewpoints. All high-resolution images (1600x1200 pixels), rendered using Mitsuba, along with associated data including camera positions, depth, 3D object models, and object masks, are stored in the standard Blender format, with support for conversion to the COLMAP format, facilitating usability across the research community.
📌 Update log
🗓️ March 2025
- Updated denoise scripts to identify and address rare missing cases caused by server-side cluster fluctuations.
- Refined benchmark results for selected algorithms (NeRO, GES, GaussianShader) on the Ablation Dataset.
- ⚠️ Note: Main benchmark results remain unaffected.
- 🔗 Updated results available at: https://christy61.github.io/openmaterial.github.io/
🗓️ November 2024
- Released benchmark results on the Ablation Dataset, with strict control over shape, material, and lighting variables.
- Benchmarked a set of representative algorithms across two tasks:
- Novel View Synthesis: Gaussian Splatting, Instant-NGP, 2DGS, PGSR, GES, GSDR, GaussianShader
- 3D Reconstruction: Instant-NeuS, NeuS2, 2DGS, PGSR, NeRO
- Updated evaluation scripts to incorporate new algorithms and support the Ablation Dataset benchmarking format.
- Improved evaluation code to better visualize benchmarking comparisons.
- 🔗 Full results available at: https://christy61.github.io/openmaterial.github.io/
🗓️ October 2024
- Released extended benchmark results on the Main Dataset:
- 7 Novel View Synthesis methods: Gaussian Splatting, Instant-NGP, 2DGS, PGSR, GES, GSDR, GaussianShader
- 6 3D Reconstruction methods: Instant-NeuS, NeuS2, 2DGS, PGSR, NeRO, NeRRF
- Highlighted algorithms specialized for challenging materials: NeRO, NeRRF, GSDR, GaussianShader.
- Updated evaluation scripts to incorporate new algorithms.
🗓️ September 2024
- Introduced a new Ablation Dataset for controlled analysis of 3D reconstruction and view synthesis.
- Controlled variables:
- Objects: Vase, Snail, Boat, Motor Bike, Statue
- Lighting: Indoor, Daytime Garden, Nighttime Street
- Materials: Conductor, Dielectric Plastic, Rough Conductor, Rough Dielectric, Rough Plastic, Diffuse
- Total: 105 unique scenes (5 × 3 × 7).
- 🔗 Data is now available.
🗓️ July 2024
- Dataset restructured for flexible material-type-based downloading.
- Users can now download subsets of data focusing on specific material categories (e.g., diffuse, conductor, dielectric, plastic).
- 📦 Updated download scripts included.
Data structure
The data structure of each subset is as follows:
├── name_of_object/[lighing_condition_name]-[material_type]-[material_name]
│ ├── train
│ │ ├── images
│ │ │ ├── 000000.png
│ │ │ |-- ...
│ │ └── mask
│ │ │ ├── 000000.png
│ │ │ |-- ...
│ │ └── depth
│ │ ├── 000000.png
│ │ |-- ...
│ ├── test
│ │ ├── images
│ │ │ ├── 000000.png
│ │ │ |-- ...
│ │ └── mask
│ │ │ ├── 000000.png
│ │ │ |-- ...
│ │ └── depth
│ │ ├── 000000.png
│ │ |-- ...
│ └── transformas_train.json
│ └── transformas_test.json