Cyan, As Mathew mentions, NuPIC is not ready for vision work. From my own investigation [1], the need for attentional and hierarchical processing is eventually required, but as you'll be aware there is a lot that can be done already before requiring cortical learning (even sub-cortical simulation of early cortical vision processing, gabor RFs, etc.). Processing towards SDR creation.
My simple experiments involved the OpenCV Retina class [2] and simple thresholding, via Otsu's method [3], on the magno-cellular channel, achieving pleasing saccades between salient features. Although I moved on to auditory sensory encoding and didn't look into whether 2D kalman filtering ([4], motion smoothing), WTA inhibition processing, and other bottom up techniques (such as [5]) can be introduced and assist in reducing the data dimensions. Only extracting minimal pertinent changing data for SDR encoding. Lots that could be added to nupic.vision before the arrival of hierarchy and attentional aspects. Is there a plan that describes the current vision changes and future in light of the recent repo move? Regards, Richard. https://uk.linkedin.com/in/richardcrowder 1 https://github.com/rcrowder/OpenCV-Attention-and-Saliency/blob/master/opencv-test/opencv-test.cpp 2 http://docs.opencv.org/doc/tutorials/contrib/retina_model/retina_model.html 3 https://en.wikipedia.org/wiki/Otsu's_method 4 https://www.youtube.com/watch?v=Jq8HcIar68Y 5 http://jov.arvojournals.org/article.aspx?articleid=2121371 On Tue, Jul 7, 2015 at 6:54 PM, Matthew Taylor <[email protected]> wrote: > Cyan, > > NuPIC is not really set up for vision work, or video processing. We > had an old implementation that could recognize humans in live video, > but that codebase is gone now. > > --------- > Matt Taylor > OS Community Flag-Bearer > Numenta > > > On Mon, Jun 29, 2015 at 8:40 AM, [email protected] <[email protected]> > wrote: > > Dear all, > > I am a computer vision engineer，Recently I consider using NAB for vision > > based human behavior abnormal detection,I have two questions。 > > 1: As I considered,NUPIC is Memory based intellgent algorithm, How can > it > > guarantee it's generalization performance？Which means different tall > human > > with same behavior wil obtain different score？ > > 2: I think the encoder step is really critical which can be considerd as > > feature generation step in Pattern Recognition. Do you have any idea > about > > encode the video to SDR? We can take more discussion. > > > > Best Regard. > > Cyan.M > > > > ________________________________ > > [email protected] > >
