Skip to content
ipanousis edited this page Apr 2, 2012 · 9 revisions

High


  • get a super simple matching working on GPU, 2 small size frames
    • think about frame rate, search window size and use tranposed A or B data. (done) with 33x33 search window, data upto transpose A, the only way to do it, originating from Gaussian pyramid)
    • if upto A only then write a function that the kernel will call... but maybe not to compute descriptors rather to compare two descriptors, this will save the storage reqs + the writes (done) descriptor built in a loop, for 'external' applications probably a function should be provided
    • written kernel function and surrounding C code, debugging stage (second half) next up
    • get it done!

Medium



Low


  • implement downsampling in a generic way wrt. kernel size + downsample rate
    • generic kernel width and downsample rate kernel working both for X and for Y (done)
    • memory buffers sorted out about what's going where while building the pyramid (done)
    • transpose the data gradients-wise (done)
    • transpose to daisy descriptors optimised (is functional to produce descriptors, but not in the interest of speed yet)
Clone this wiki locally