Compressed Spatial Frequency Transfrom (CSFT)

For more information, contact:
Ke Gao, kegao@mail.missouri.edu
Prof. K. Palaniappan, palaniappank@missouri.edu
226 Naka Hall
University of Missouri-Columbia
Columbia, MO 65211

Description

CSFT is novel feature detection and description approach for robustly matching features across different views suitable for applications such as camera pose estimation and Structure-from-Motion in large scale aerial images. CSFT features are robust across a range of scale and rotation, image blur, varying illumination, and JPEG compression artifacts. The CSFT descriptor is compressed using geometric transformations in the frequency domain that significantly reduce its size. The CSFT feature uses a novel vectorized Normalized Cross Correlation (VNCC) method for matching descriptors. CSFT is a domain independent method that can be easily implemented. CSFT requires no training step compared to the deep learning approaches.

Building CSFT

Build Library Dependencies

Before CSFT can be built, several dependencies need to be available on the system.

OpenCV C++ >= 3.4.0
CMake
OpenMP
CUDA (optional)

CUDA is optional but recommended for a better run-time performance. CMake flag -DNO_CUDA_BUILD can be used to adjust the build to switch between GPU and CPU implementation as needed. CSFT is built without CUDA when set to -DNO_CUDA_BUILD=1. Otherwise, CMake defaults to building with CUDA support when not explicitly specified, i.e., -DNO_CUDA_BUILD=0.

Configure and Build

$ cd CSFT
$ mkdir build
$ cd build
$ cmake ..
$ make

Usage

Running CSFT

Once CSFT has been successfully built, there should be a executable in the bin directory: CSFT.

To run CSFT, go to ./bin/linux for Linux or ./bin/macos for MacOS, and invoke ./CSFT <parameters_config_file>.
Example usage: ./CSFT ../../params.txt

Description of the CSFT Parameters in Config File

img1_fullpath: path to the 1st input image
img2_fullpath: path to the 2nd input image
out_path: path to the output folder
n_keypoints: number of keypoints to detect per image
upright_csft: flag for using upright-CSFT which disables rotation invariance (1: upright-CSFT, 0: standard-CSFT)
flag_export_keypts: flag for exporting feature keypoints to text files
flag_export_desc: flag for exporting feature descriptors to text files
flag_export_match: flag for exporting feature matches to text files
flag_draw_keypts_img: flag for visualizing feature keypoints on the input images
flag_draw_match_img: flag for visualizing feature matches on the input images

Output

The output files will be saved to the directory specified by out_path in the parameters config file.

Visualization

Images showing the detected keypoints will be saved if flag_draw_keypts_img is set to 1.
An image showing the matches will be saved if flag_draw_match_img is set to 1.

Text Files

If flag_export_keypts is set to 1, an ASCII text file containing the keypoints locations (x y) will be created for each input image where each row corresponds to a keypoint.
Example:

618.116028 204.111755
437.704346 299.728333
597.402649 202.520157
188.808380 446.749207
437.238708 306.217590

If flag_export_desc is set to 1, an ASCII text file containing the descriptors will be created for each input image where each row corresponds to the descriptor for a keypoint.
If flag_export_match is set to 1, an ASCII text file containing the matches (x1 y1 x2 y2) will be created for a pair of images where each row corresponds to a match. x1 y1 and x2 y2 denote the coordinate of the matched keypoint in the first image and second image, respectively.
Example:

421.005707 304.459961 406.880188 321.907928
409.846680 297.340576 397.476959 325.259186
392.506470 306.396332 393.034668 338.806610
376.126678 328.696777 392.975586 357.134857
437.796295 402.756653 462.513184 370.383728