Implementation of 2D/3D geodesic distance transforms
Geodesic transformation of images can be implementated with two approaches: fast marching and raster scan. Fast marching is based on the iterative propagation of a pixel front with velocity F [1]. Raster scan is based on kernel operations that are sequentially applied over the image in multiple passes [2]. In GeoS [3], the authors proposed to use a 3x3 kernel for forward and backward passes for efficient geodesic distance transform, which was used for image segmentation.
Raster scan for geodesic distance transform. Image from [3].
DeepIGeoS [4] proposed to combine geodesic distance transforms with convolutional neural networks for efficient interactive segmentation of 2D and 3D images.
- [1] Sethian, James A. "Fast marching methods." SIAM review 41, no. 2 (1999): 199-235.
- [2] Borgefors, Gunilla. "Distance transformations in digital images." CVPR, 1986
- [3] Criminisi, Antonio, Toby Sharp, and Andrew Blake. "Geos: Geodesic image segmentation." ECCV, 2008.
- [4] Wang, Guotai, et al. "Deepigeos: A deep interactive geodesic framework for medical image segmentation." arXiv preprint arXiv:1707.00652 (2017). https://arxiv.org/abs/1707.00652
A comparison of fast marching and ranster scan for 2D geodesic distance transform.
This repository provides a cpp implementation of fast marching and raster scan for 2D/3D geodesic distance transforms, and proivdes a python interface to use it. This is part of the work of DeepIGeoS [4]. If you use our code, please cite this paper. https://arxiv.org/abs/1707.00652
How to use
1, build the cpp files with python, run python setup.py build
2, install the package to default python path, run python setup.py install
3, see a 2D example, run python test2d.py
4, see a 3D example, run python test3d.py