3DFRONT-NC is an enhanced dataset derived from 3D-FUTURE and 3D-FRONT. This dataset introduces a cuboid-based structural representation of furniture and utilizes it to refine 3D indoor scene layouts. The process consists of two main stages:

We convert complex 3D-FUTURE meshes into cuboid assemblies to represent furniture models, i.e., a set of axis-aligned cuboids approximating the furniture's geometry and structure. The resulting cuboid assemblies offer compact, interpretable, and editable representations of furniture. We have processed a total of 4,356 furniture pieces used in 3D-FRONT and stored them in the part1/data/cuboids_dict.json file.

To visualize the cuboid assemblies, you can use the provided script to export them to OBJ format. The exported model size should match the size of the corresponding raw_model.obj in 3D-FUTURE.

# Navigate to the part1/scripts directory
cd part1

# Export a specific model's cuboid assembly to OBJ format
python scripts/export_obj.py --sample_id 1c648cfd-6eca-4a84-af58-b62a3a4d173a

This will create an OBJ file named output_1c648cfd-6eca-4a84-af58-b62a3a4d173a.obj that can be opened in any 3D modeling software.

We use the structured cuboid assemblies to enhance the original 3D-FRONT indoor scenes. We adjust object placements to resolve collisions, enhance support relationships, and ensure stability.

Note: Although the scenes are significantly improved, some collision issues may still remain in the dataset.

• Release Part 1 data (cuboid assemblies for 3D-FUTURE models)
• Release Part 2 data (refined 3D-FRONT scenes)
• Release the code for generating cuboid assemblies
• Release the code for refining 3D-FRONT scenes using cuboid structures