A multi-thread CPU implementation of 3D SIFT, the Scale invariant feature transform (SIFT) for 3D image (volumetric image). The feature matching of 3D SIFT features is also provided.

The API of reading NIFTI images is included in this program.

The program is written in C++ language and is parallelized using OpenMp.

• Read nifti image:
  • Header file:Include/Util/readNii.h
  • Function:readNiiFile(path, nx, ny, nz);
  • The function read the path, and return a pointer with the data of volume image, and the dimensions of the image are written on nx, ny and nz .
• 3D SIFT feature extraction.
  • Header file:Include/cSIFT3D.h
  • Object create function:CPUSIFT::CSIFT3DFactory::CreateCSIFT3D(volData, nx, ny, nz)
    • This is a factory function belongs to namespace CPUSIFT, it creates a 3D SIFT object with the datapointer volData and dimensions nx,ny and nz as input. Also, the parameters of 3D SIFT can be indicated, they are set as default values, as seen in the declaration of the function.
    • The created 3D SIFT object should be deleted.
  • Extract class function: CSIFT3D->KpSiftAlgorithm();
    • perform feature extraction procedure
  • Obtain keypoint function:CSIFT3D->GetKeypoints()
    • return vector of extracted keypoints.
• Feature matching on two sets of features.
  • Header file:Include/cMatcher.h
  • Class: CPUSIFT::muBruteMatcher
  • This class provides several functions of brute-force matching.
    • Overall: The enhancedMatch is recommended.
    • Parameters: (matchRefCoor, matchTarCoor, vRefKp, vTarKp, thrshold)
      • The first two parameters are coordinates of returned matched keypoints, representing coordinates in reference image and target image.
      • vRefKp and vTarKp represents the two sets of keypoints extracted from the image
      • threshold is the threshold for filtering ambiguous matches. The smaller the value, the more strict the filtering. 0.85 is recommended and used as default parameter.
    • injectMatch(): perform feature matching from reference set of keypoint to target set of keypoint. Like mostly used in 2D SIFT.
    • bijectMatch(): perform matching from reference to target, and also from target to reference, the valid matches must be appeared in the both two matching, as used in Ref[1].
    • enhancedMatch(): perform match from the reference to target first, and then for target features that are matched by multiple reference feature the matching from target to reference is performed to remove ambiguity, as used in Ref[2].

Currently, the visual studio project is provided, it's recommended to use windows PC to run the codes.

A example of feature extraction and matching is provided in example.cpp, which is included in the visual studio project too. The steps of running the example data is as follows:

1. Open the visual studio project .sln file.
2. Set the Windows SDK version of the visual studio project to that are installed on your PC. PS: version below 10.0.17763.0 is non-tested.
3. Modify the path of data files in example.cpp
4. Build and run the codes

If the readers want to use this library in another programs, two methods are recommended:

1. Put all of the source codes of this project (besides example.cpp) into the another program, and the dependencies of different files and third-party libraries should be handled correctly. An example can be seen in our another project 3D SIFT PiDVC.
2. Build dll and lib files using the visual studio projects by setting the Configuration Types of visual studio project to be Dynamic Library (.dll) instead of Application (.exe). And in the program importing 3D SIFT, it's required to include only the header files of 3D SIFT and import the .lib flies for compilation, and the .dll files are required for running the generated programs.

Several nifti files are provided:

https://drive.google.com/open?id=1IlOBp1hu-KUh648lXR3YjgNih71Mj8ke

Several third-party libraries are used in this work, and are also included in this project.

• Eigen, put in path 3DSIFT\3party\Eigen, used to perform matrix computation, such as eigendecompostion.
• layNii, put in path 3DSIFT\3party\layNii, used to read Nifti format files.
• Zlib, put in path 3DSIFT\3party\zlib, used by layNii. It should be noticed that the provided lib files are built on x64 Windows 10 with Visual studio 2017 and WINDOWS SDK 10.0.17763.0 from the source code.

This program is based on the work of Rister. He has proposed a well-performed 3D SIFT with rotation-invariance, as seen in Reference [1].

There are several different details in our implement:

• The Gaussian Smoothing is fixed in this work, by removing the zooming of gaussian kernel in higher octave. A series of gaussian kernels are used in different octaves without zooming of kernel radius. The produced gaussian images are verified by comparing with that produced by matlab with the same sigma parameter.
• In descriptor construction, the keypoint is located at the center of the sub-regions as described in the paper (reference [1]) , while in this implementation the keypoint is coincide with one of the sub-region center.

If you want to cite this work, please refer to the papers as follows:

[1] B. Rister, M. A. Horowitz and D. L. Rubin, "Volumetric Image Registration From Invariant Keypoints," in IEEE Transactions on Image Processing, vol. 26, no. 10, pp. 4900-4910, Oct. 2017. doi: 10.1109/TIP.2017.2722689

[2] Our paper in the manuscript.