I have customerize intel_ocl_motion_estimation_advanced to dump motion vectors in to a text file and used it for bench-marking on KITTI tool. KITTI is used for optical flow motion vector bench-marking which can be accessed from : http://www.cvlibs.net/datasets/kitti/eval_scene_flow.php?benchmark=flow. The input to KITTI is a png file so i converted motion vectors per 8x8 to a pixel based png file where each 8x8 block has same motion vector.
KITTI provide its own set of optical flow results(ground truth). I used intel ME results against KITTI ground truth(GT) for 194 clips. The disparity between GT and Intel ME is 32% which looks quite high in number. I figured out intel ME only works for even resolution so i black padded the input file after that i got 24% disparity for block 8x8 and 26% for block 4x4(again surprising). But this is the best result so far.
Command line: MotionEstimationAdvanced.exe --input ./outputyuv/000000_11.yuv --output ./intel_8x8/intel_8x8_0.txt --width 1226 --height 370 --precision quarter --block 8x8
Can i improve better here through command line?
Is their a way to turn off intra MB ?
Is their a way to feed external Hints ?