Tengpaz/worldrenderer-dataset-test

WorldRenderer Dataset Test

Dataset Summary

WorldRenderer Dataset Test is a synthetic multi-scene 3D rendering dataset designed for research in:

• Novel View Synthesis (NVS)
• Neural Rendering
• Geometry-aware Generation
• Multi-view Representation Learning
• World Models
• 3D-conditioned Generative Modeling

The dataset contains 27 textured 3D scenes.
Each scene is rendered using a predefined monocular camera trajectory consisting of 401 frames. For every frame, aligned multi-modal rendering outputs are provided, including:

• RGB images
• Depth maps
• Surface normal maps

Additionally, each scene also contains:

• the original textured 3D scene asset
• the first rendered frame
• an automatically generated caption for the first frame

This dataset is intended to provide aligned geometric and visual supervision for modern multi-view generation and rendering systems.

---

Dataset Characteristics

Property	Value
Number of scenes	27
Frames per scene	401
Camera trajectory	Single monocular trajectory
Modalities	RGB / Depth / Normal
Geometry assets	Included
Rendering type	Synthetic
Alignment	Pixel-aligned multi-modal

---

Supported Tasks

This dataset can be used for:

• Novel View Synthesis
• Neural Rendering
• Multi-view Diffusion Models
• Geometry-aware Image Generation
• 3D Reconstruction
• Surface Normal Estimation
• Depth Prediction
• Camera-conditioned Generation
• World Modeling

---

Dataset Structure

Each scene is stored in an independent folder. Example directory structure:

worldrenderer-dataset-test/
├── 0000/
│   ├── rgb/
│   │   ├── rgb_000000.png
│   │   ├── rgb_000001.png
│   │   └── ...
│   │
│   ├── depth/
│   │   ├── depth_000000.png
│   │   ├── depth_000001.png
│   │   └── ...
│   │
│   ├── normal/
│   │   ├── normal_000000.png
│   │   ├── normal_000001.png
│   │   └── ...
│   │
│   ├── first_frame.png
│   ├── model.blend
│
├── 0001/
│   └── ...
│
├── ...
│
└── model.tar.gz

⸻

Naming Convention

All frame indices start from 0.

Frame naming format:

rgb_000000.png depth_000000.png normal_000000.png

The final frame index is:

000400

corresponding to a total of 401 frames per scene.

⸻

Data Fields

RGB

Rendered RGB images along the camera trajectory.

Example:

rgb/rgb_000123.png

⸻

Depth

Depth maps aligned with RGB images.

Example:

depth/depth_000123.png

Depth values are rendered directly from the 3D scene geometry.

1mm metric

65.535m max depth

⸻

Normal

Surface normal maps aligned with RGB images.

Example:

normal/normal_000123.png

Normals are represented in camera space.

⸻

Model

Original textured 3D scene asset.

model/

⸻

Data Generation Pipeline

For each scene:

1. A textured 3D environment is loaded.
2. A monocular camera trajectory is generated.
3. 401 aligned frames are rendered.
4. RGB, depth, and normal maps are exported.
5. All assets are organized into scene-wise folders.

⸻

Intended Uses

The dataset is designed for research purposes, including:

• training neural rendering systems
• studying geometry-aware generation
• evaluating multi-view consistency
• camera-conditioned generation
• 3D scene understanding
• synthetic world modeling

Potential model families include:

• NeRF-based methods
• Gaussian Splatting pipelines
• Diffusion Transformers
• Multi-view autoregressive models
• Geometry-aware diffusion models
• Vision-language world models

⸻

Example Usage

Basic File Access

from pathlib import Path root = Path("worldrenderer-dataset-test") scene_dir = root / "0000" rgb_path = scene_dir / "rgb" / "rgb_000000.png" depth_path = scene_dir / "depth" / "depth_000000.png" normal_path = scene_dir / "normal" / "normal_000000.png" print(rgb_path)

⸻

Recommended Research Directions

This dataset is particularly suitable for:

• M-to-N view generation
• Geometry-conditioned diffusion
• Camera-conditioned transformers
• Unified rendering and reconstruction
• World representation learning
• Multi-modal scene understanding

⸻

Limitations

• Synthetic rendering only
• Limited scene diversity
• Single trajectory per scene
• Automatically generated captions may contain inaccuracies
• Rendering configuration may differ from real-world camera distributions

⸻

Contact

For questions, issues, or collaboration opportunities, please open an issue on the dataset repository page.