This is my implementation of the ECC image alignment algorithm for GPU, using CUDA and NPP (NVIDIA 2D Image And Signal Performance Primitives) libraies.
As far as I concern, this is the only optimization available over OpenCV's CPU implemenation.
The ECC Finds the geometric transform (warp) between two images in terms of the ECC criterion.
Georgios D Evangelidis and Emmanouil Z Psarakis. Parametric image alignment using enhanced correlation coefficient maximization.
Pattern Analysis and Machine Intelligence, IEEE Transactions on, 30(10):1858–1865, 2008.
- Supports only 1-Channel images (grayscale). Implementation for 3-Channel is trivial.
- CMakeLists file might not suffice for linking library dependencies (OpenCV and NPP) and folder includes, add it by yourself.
- Opencv compiled with CUDA 10.2
- CUDA Toolkit (has the NVIDIA NPP libaries)
Benchmark was conducted using the following command argument (the input image is obtained automatically from opencv installation directory):
./ecc_gpu fruits.jpg -o=outWarp.ecc -m=homography -e=1e-6 -N=70 -v=1
On 1920x2560 (rowsxcols) image resolution:
| Device | Time [sec] | Implementation |
|---|---|---|
| Single-Thread CPU (Intel i7-6700) | 20.865 | OpenCV CPU (Using AVX2) |
| GPU (Nvidia GTX 970 - NO OC) | 2.187 | OpenCV GPU+NPP |
Order of magnitude (x10~) speed-up !!
Expect better speed-up with better GPU and harder alignment cases (bad alignment initilization, more iterations, small epsilon).