Improved confusion matrix generation, so it uses samples that are roughly uniformly distributed between the 2 classes, across the entire validation set. The matrix agrees with the reported accuracy, suggesting something is wrong with my implementation used with OpenCV in c++ (as the acc. is lower).