TILDE: a Temporally Invariant Learned DEtector

This software is the C++ and MATLAB implementation for the TILDE keypoint detector presented in [1]. The software provides both the C++ implementation of TILDE which can be used easily to detect keypoints, and a MATLAB based evaluation framework.

This software is strictly for academic purposes only. For other purposes, please contact us. When using this software, please cite [1] and other appropriate publications if necessary (see matlab/external/licenses for details).

[1] Y. Verdie, K. M. Yi, P. Fua, and V. Lepetit. "TILDE: A Temporally Invariant Learned DEtector.", Computer Vision and Patern Recognition (CVPR), 2015 IEEE Conference on.

Contact:

Yannick Verdie : yannickverdieepflch Kwang Moo Yi : kwangyiepflch

            << The TILDE C++ Implementation >>

The C++ implementation of TILDE provides an easy-to-use library with a simple demo program to detect and display the keypoints. We provide TILDE keypoints learned with the Webcam dataset in [1].

NOTES:

• Code compiles with clang (macOS) and gcc-4.9 (Linux)
• The code will generate a static library, a dynamic library (libTILDE) and an example code (demo).

REQUIREMENTS:

• OpenCV 2.4.9 or higher
• CMake 2.8 or higher

USAGE:

Build the libraries and the demo. Standard procedure is as follow (from the project root directory):

cd c++ mkdir build cd build cmake .. make

Then you can run the demo code from the build directory with:

./Demo/demo

IMPORTANT FUNCTIONS:

The main function is getTILDEKeyPoints (see demo.cpp)

std::vector getTILDEKeyPoints(cv::Mat image, std::string pathFilter, bool useApprox, bool sortKeypoints, bool keepPositiveOnly, cv::Mat *score)

<>:

• image: a openCV U8C3 Mat object representing the image to process
• pathFilter: a std string object giving the name of the filter to apply
• useApprox: a boolean indicating to use TILDE (false) or the approximated version of TILDE (true).
• sortKeypoints: (true) we sort the keypoints by decreasing order of their scores
• keepPositiveOnly: (true) only the keypoints with positive score are returned
• score: a pointer to an openCV Mat image, if the pointer is not null, the score map is retuned in addition to the keypoints

<

>:

• std::vector: a std vector of OpenCV KeyPoint object (see OpenCV doc for more details on how to use them)

DIRECTORY STRUCTURE:

              When  a number  is added  in the  name of  the filter,  it
          denotes that the filter is  for use with the approximation
          flag  on (i.e.  it is  the approximated  TILDE). The  name
          indicates which dataset was used to learn this filter.