CrossSDF: 3D Reconstruction of Thin Structures from Cross-Sections

CrossSDF is a novel method for 3D reconstruction of thin structures from cross-sectional data. This repository contains the dataset used for evaluation in our CVPR 2025 paper.

🎬 Results & Demos

Visit our project website to see:

• Interactive video comparisons of CrossSDF vs. other methods
• Method overview and technical details
• Training progress visualization
• Qualitative results on medical structures

📋 Table of Contents

• Results & Demos
• Dataset Overview
• Data Format
• Download and Usage
• Benchmarking
• Citation
• License
• Contact

📊 Dataset Overview

Our dataset consists of 18 high-quality 3D meshes organized into two categories for comprehensive evaluation of 3D reconstruction methods on both thin and thick structures.

Thin Structures Dataset

The thin structures dataset contains 6 meshes featuring complex tubular and branching geometries, ideal for evaluating reconstruction of fine anatomical networks:

Mesh	Description	Source
heart.stl	Complete cardiac mesh	Vascular Model Repository
pulmonary_vascular_tree.stl	Full branching lung vasculature	VMR
pulmonary_arteries.stl	Major pulmonary arterial segments	VMR
cerebral_arteries.stl	Circle of Willis and cerebral vasculature	VMR
coronary_arteries_v1.stl	Coronary artery topology (variant 1)	VMR
coronary_arteries_v2.stl	Coronary artery topology (variant 2)	VMR

Thick Structures Dataset

The thick structures dataset includes 6 meshes for evaluating generalizability beyond vascular structures:

Mesh	Description	Source
armadillo.stl	Armadillo model	OReX Dataset
brain.stl	Brain surface mesh	OReX Dataset
eight.stl	Figure-eight topology	OReX Dataset
hand.stl	Human hand model	OReX Dataset
mammoth.stl	Mammoth model	OReX Dataset
ok.stl	OK hand gesture	OReX Dataset

Additional Real-World Datasets

For validation on real medical data, we also utilize:

• IRCADb-01: 20 contrast-enhanced abdominal CT scans with manual liver vascular segmentations
• Medical Segmentation Decathlon: Task 08 with 443 portal-phase CT scans and hepatic vessel annotations

🔧 Data Format

All meshes are provided in STL format with the following specifications:

• Format: Binary STL files
• Coordinate System: All meshes are normalized to the range [-1, 1] in all three dimensions
• Units: Dimensionless (normalized coordinates)
• Orientation: Consistent across all models
• Quality: High-resolution meshes suitable for research and benchmarking

File Structure

cross_sdf/
├── data/
│   ├── thin_structures/          # Thin/tubular geometries
│   │   ├── heart.stl
│   │   ├── pulmonary_vascular_tree.stl
│   │   ├── pulmonary_arteries.stl
│   │   ├── cerebral_arteries.stl
│   │   ├── coronary_arteries_v1.stl
│   │   └── coronary_arteries_v2.stl
│   └── thick_structures/         # Thick/solid geometries
│       ├── armadillo.stl
│       ├── brain.stl
│       ├── eight.stl
│       ├── hand.stl
│       ├── mammoth.stl
│       └── ok.stl
├── assets/
│   ├── images/                   # Figures and visualizations
│   └── videos/                   # Comparison videos
├── README.md
├── LICENSE
└── CITATION.cff

📥 Download and Usage

Clone Repository

git clone https://github.com/iamsalvatore/cross_sdf.git
cd cross_sdf

Loading STL Files

The meshes can be loaded using any standard 3D processing library:

Python (using trimesh):

import trimesh

# Load a thin structure
mesh = trimesh.load('data/thin_structures/heart.stl')
print(f"Vertices: {len(mesh.vertices)}, Faces: {len(mesh.faces)}")

Python (using Open3D):

import open3d as o3d

# Load a thick structure
mesh = o3d.io.read_triangle_mesh('data/thick_structures/armadillo.stl')
mesh.compute_vertex_normals()
o3d.visualization.draw_geometries([mesh])

MATLAB:

% Load mesh
mesh = stlread('data/thin_structures/pulmonary_vascular_tree.stl');

🏆 Benchmarking

This dataset is designed for evaluating 3D reconstruction methods, particularly those dealing with:

• Cross-sectional reconstruction from parallel or random planar cuts
• Thin structure preservation in tubular and branching geometries
• Multi-scale geometry handling from fine vessels to solid objects
• Robustness to noise and incomplete data

Evaluation Metrics

Common metrics for benchmarking include:

• Chamfer Distance (CD)
• Hausdorff Distance
• Normal Consistency
• Volume Preservation
• Surface Quality Metrics

📖 Citation

If you use this dataset in your research, please cite our paper:

@inproceedings{walker2025_crosssdf,
    author = {Walker, Thomas and Esposito, Salvatore and Rebain, Daniel and Vaxman, Amir and Onken, Arno and Li, Changjian and Mac Aodha, Oisin},
    title = {CrossSDF: 3D Reconstruction of Thin Structures from Cross-Sections},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    year = {2025},
}

📄 License

This dataset is released under a custom license for non-commercial research use only. See LICENSE for full details.

📧 Contact

For questions about the dataset or paper, please contact:

• Salvatore Esposito - salvatore.esp95@gmail.com
• Thomas Walker - thomas.walker@ed.ac.uk

🙏 Acknowledgments

We thank the creators of the Vascular Model Repository and the OReX Dataset for making their high-quality 3D models publicly available.

---