This project has been created by Vito Walter Anelli and Alessandro Pagliara.
The OpenCV matchTemplate primitive for gpu has been adapted in order to enable calculations on 32-bit floating-point raw data.
The detection of the offsets that maximize the crosscorrelation between master and slave oversampled patches is a time-consuming task that typically must be repeated thousands times for each interferometric acquisition pair: it has been hence selected in this study as a good starting point for providing a preliminary assessment of GPU performances with respect to single-thread or multi-thread CPU implementations.
OpenCV libraries ver.2.4.9 are used for oversampling patches and cross-matching them in both CPU or GPU environment. The normalized cross-correlation is computed in the frequency domain (FFT) with the matchTemplate primitive, whose source code has been adapted in order to enable calculations on 32-bit floating-point raw data. Concerning the GPU implementation, all GPU cores are involved in the oversampling and matching of a single pair of master and slave patches; the process is sequentially iterated for all the selected patches in input.
If you are interested in it please contact us:
Vito Walter Anelli - vitowalter.anelli@poliba.it
Research Lab page
linkedin
Alessandro Pagliara
linkedin
Feel free to use or extend our code, just cite this:
Efficient implementation of InSAR time-consuming algorithm kernels on GPU environment
@inproceedings{guerriero2015efficient,
title={Efficient implementation of InSAR time-consuming algorithm kernels on GPU environment},
author={Guerriero, Andrea and Anelli, Vito Walter and Pagliara, Alessandro and Nutricato, Raffaele and Nitti, Davide Oscar},
booktitle={Geoscience and Remote Sensing Symposium (IGARSS), 2015 IEEE International},
pages={4264--4267},
year={2015},
organization={IEEE}
}