canondetortugas
commented
11 years ago Also, maybe we should just not use the OpenCV SVM library. We could give libSVM a shot or just not use SVMs at all, since it seems like the SVM approach tends to simplify to segmenting off part of the hue spectrum.
The SVM predict function from OpenCV's ml module, which we use to classify color images, is very slow. With radial basis functions, in the simplest case where we only use two support vectors, the function takes around 500ms to classify a full 640x480 image. Since each pixel is computed independently, it would be trivial to parallelize this operation if we had a GPU. According to the OpenCV docs, the function is parallelized with the Intel TBB library. We should see if we are currently compiling with this option or figure out how to do so if we aren't. Also, since the shapes we are trying to detect are very primitive, we could get away with downsampling the images before we classify them. Even a scaling them by 1/2 would speed things up significantly.