SceneMaker: Open-set 3D Scene Generation with Decoupled De-occlusion and Pose Estimation Model

Yukai Shi^1,3, Weiyu Li^2,4, Zihao Wang⁴, Hongyang Li³, Xingyu Chen³, Ping Tan^2,4, Lei Zhang³

¹ Tsinghua University    ² HKUST    ³ IDEA Research    ⁴ LightIllusions

SceneMaker dataset is a 3D scene dataset for open-world scenarios from paper SceneMaker. This dataset is based on the Objaverse dataset, where 3D objects are randomly selected and placed in a physically plausible manner to construct 3D scenes. Each scene contains multi-view camera intrinsics and extrinsics, along with rendered images including RGB, depth maps, and segmentation maps.

📖 Dataset Description

Each scene (identified by UUID) contains the following elements:

• Main Objects: 2-5 original objects from Objaverse, represented by 20,480 point clouds extracted from meshes with corresponding normal directions
• Environment Elements:
  • Ground plane
  • Environment map
• Number of Views: Each scene contains 20 different viewpoints
• Rendering Types: Each viewpoint provides three types of images:
  • RGB: Color-rendered scene images
  • Depth: Scene depth maps
  • Seg: Segmentation maps for each object

Sample scenes from the SceneMaker dataset showing RGB images, depth maps, and segmentation masks

📁 Dataset Structure

After extraction, the dataset has the following structure:

dataset/
├── 000-001.tar.gz to 000-159.tar.gz    # Split archives containing scene render results
├── geometry/                            # Folder containing object geometry data
│   └── geometry.tar.part_*             # Split archives of 3D object meshes/point clouds
│
└── [After extraction]
    ├── 000-000/                        # Folder for archive 000-000
    │   ├── {scene_uuid_1}.json         # Scene configuration (object count, IDs, poses, scene size)
    │   ├── {scene_uuid_2}.json         # Scene configuration for another scene
    │   ├── ...
    │   ├── {scene_uuid_1}/             # Scene folder (identified by UUID)
    │   │   ├── 0_rgb.png               # RGB render from view 0
    │   │   ├── 0_depth.png             # Depth map from view 0
    │   │   ├── 0_seg.png               # Segmentation map from view 0
    │   │   ├── 1_rgb.png               # RGB render from view 1
    │   │   ├── 1_depth.png             # Depth map from view 1
    │   │   ├── 1_seg.png               # Segmentation map from view 1
    │   │   ├── ...                     # Views 2-18
    │   │   ├── 19_rgb.png              # RGB render from view 19
    │   │   ├── 19_depth.png            # Depth map from view 19
    │   │   ├── 19_seg.png              # Segmentation map from view 19
    │   │   └── metadata.json           # Rendering information for each view (camera intrinsics/extrinsics)
    │   │
    │   ├── {scene_uuid_2}/             # Another scene in 000-000
    │   │   └── ...
    │   └── ...
    │
    ├── 000-001/                        # Similar structure as 000-000
    ├── ...                             # Folders 000-002 to 000-158
    ├── 000-159/                        # Similar structure as 000-000
    │
    └── geometry/                        # Extracted object geometries
        ├── geometry.json               # Index of all object UUIDs and their file paths
        └── sample/                      # Object point clouds with normals in npz format
            ├── 000-000/                 # Objects from scenes in 000-000
            │   ├── {object_uuid_1}.npz  # Point cloud and normals for object 1
            │   ├── {object_uuid_2}.npz  # Point cloud and normals for object 2
            │   └── ...
            ├── 000-001/                 # Similar structure
            ├── ...                      # Folders 000-002 to 000-158
            └── 000-159/                 # Similar structure

📋 File Descriptions

• 📝 Scene Configuration Files ({scene_uuid}.json): JSON files in each numbered folder containing:
  • Object count: Number of objects in the scene
  • Object IDs: List of object UUID identifiers
  • Object poses: Translation, rotation, and scale for each object
  • Scene size: Bounding box dimensions of the scene
• 📂 Scene Folders ({scene_uuid}/): Each folder represents one scene with a unique UUID identifier
• 🖼️ RGB Images (*_rgb.png): Color-rendered images of the scene from different viewpoints (512×512)
• 📏 Depth Maps (*_depth.png): Depth information for each pixel with range 0-20, stored as 16-bit PNG
• 🎨 Segmentation Maps (*_seg.png): Instance segmentation masks where each object is represented by an integer starting from 0, following the order of objects in the corresponding {scene_uuid}.json file (1 is plane)
• 📊 Metadata (metadata.json): Rendering information for each view, including camera intrinsics and extrinsics for all 20 views
• 🗂️ Geometry Index (geometry/geometry.json): Index file mapping all object UUIDs to their corresponding file paths
• ☁️ Object Point Clouds (geometry/sample/*/*.npz): NPZ files containing object point clouds (20,480 points × 6 dimensions), where the first 3 dimensions are XYZ coordinates and the last 3 dimensions are normal vectors, organized by archive number

🚀 Usage Instructions

Extract the render results of scenes

tar -xzvf 000-*.tar.gz

Extract the object geometry in the scenes

cd geometry
cat geometry.tar.part_* | tar -xv

Or merge them first, then extract:

cd geometry
cat geometry.tar.part_* > geometry.tar
tar -xvf geometry.tar
rm geometry.tar  # Optional: remove the merged file after extraction

File Naming Convention

Each scene in the dataset is named by UUID and contains multi-view rendering results. Both scenes and objects are uniquely identified by their UUID values. When using this dataset, please update the file paths to match your dataset root directory.

📚 Citation

If you use this dataset, please cite the SceneMaker and Objaverse datasets:

@article{shi2025scenemaker,
  title={SceneMaker: Open-set 3D Scene Generation with Decoupled De-occlusion and Pose Estimation Model},
  author={Shi, Yukai and Li, Weiyu and Wang, Zihao and Li, Hongyang and Chen, Xingyu and Tan, Ping and Zhang, Lei},
  journal={arXiv preprint arXiv:2512.10957},
  year={2025}
}

@article{objaverse2023,
  title={Objaverse: A Universe of Annotated 3D Objects},
  author={Deitke, Matt and Schwenk, Dustin and Salvador, Jordi and Weihs, Luca and Michel, Oscar and VanderBilt, Eli and Schmidt, Ludwig and Ehsani, Kiana and Farhadi, Ali and Kembhavi, Aniruddha and others},
  journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2023}
}