webxos/underworld_dataset_v2

 _   _ _   _______ ___________ _    _  ___________ _    ______ 
| | | | \ | |  _  \  ___| ___ \ |  | ||  _  | ___ \ |   |  _  \
| | | |  \| | | | | |__ | |_/ / |  | || | | | |_/ / |   | | | |
| | | | . ` | | | |  __||    /| |/\| || | | |    /| |   | | | |
| |_| | |\  | |/ /| |___| |\ \\  /\  /\ \_/ / |\ \| |___| |/ / 
 \___/\_| \_/___/ \____/\_| \_|\/  \/  \___/\_| \_\_____/___/  
      

Underworld Dataset v2

Generated with WEBXOS UNDERWORLD LANDSCAPE GENERATOR

Download the GENERATOR from the /underworld/ folder of this repo to create your own datasets.

Optimized for Hugging Face Overworld model training

Dataset Details

• Frames: 120
• Resolution: 1280x562
• Objects: 150
• Terrain Size: 160x160
• View Mode: wireframe

Usage for Training

1. Extract the dataset
2. Use frames/ folder for image sequences
3. Use metadata.json for generation parameters
4. Use labels.csv for object detection training
5. Use with Hugging Face transformers or custom training scripts

WEBXOS Underworld Dataset v2

WEBXOS Underworld Dataset v2 is a large-scale synthetic wireframe landscape dataset designed for training and evaluating computer vision models, particularly for 3D reconstruction, depth estimation, and geometric understanding tasks. Generated using a custom Three.js-based procedural generator, this dataset provides clean, structured data with perfect ground truth annotations.

Key Features:

• Multi-modal Annotations: RGB images, depth maps, surface normals, semantic segmentation, instance segmentation
• Perfect Ground Truth: No annotation noise, perfect correspondence across modalities
• Procedural Diversity: Parametrically generated landscapes with controlled variations
• Camera Trajectories: Multi-view sequences for 3D reconstruction and novel view synthesis
• Wireframe Focus: Emphasis on geometric structures and edges for edge detection and line segment detection tasks

Primary Tasks:

• Depth Estimation: Predict depth from single RGB images
• 3D Reconstruction: Reconstruct 3D scenes from single or multiple views
• Semantic Segmentation: Classify terrain and object categories
• Instance Segmentation: Identify individual object instances
• Surface Normal Estimation: Predict surface orientations
• Edge Detection: Detect wireframe edges and geometric boundaries
• Novel View Synthesis: Generate new views of 3D scenes
• Object Detection: Detect and localize objects in 3D space

Data Fields

• image: RGB wireframe render (512×512 or 1024×1024)
• depth: Ground truth depth map (float32, normalized 0-1)
• normal: Surface normal map (3-channel float32)
• segmentation: Semantic segmentation mask (8 classes)
• instances: Instance segmentation mask
• bounding_boxes: 3D bounding boxes with categories
• camera_params: Camera intrinsics and extrinsics
• terrain_params: Procedural generation parameters
• object_metadata: Object positions, rotations, scales

Data Splits

• train: 50,000 scenes (40,000 unique terrains + 10,000 augmentations)
• validation: 5,000 scenes
• test: 5,000 scenes (held-out for benchmarking)

Citation

If you use this dataset, please cite: WEBXOS UNDERWORLD Landscape Generator, 2026

License

cc-by-4.0