I am looking to use GPU support in Opencv for my project. Currently, OpenCV provides OpenCL and CUDA support which is great for a majority of the functions i need. However, i have a pretty exotic match template function. Each point in the template is warped by a given homography and the SQ_DIFF_NORM is calculated for that point in the template and added to the result.
The method needs to do the following.
For an image patch in the template, warp each point using perspective transform with a given homography and perform the square difference between the source and target images. Match template, point by point for specifically warped points.
Known search ROI.
Known template ROI.
Find best match point in search ROI which is warped by homography H. Each template point needs to be warped by the homography and matched against the search image location. Interpolation should be required for the warped point.
How much can this be accelerated in OpenCV using the GPU?