Optical flow for fine-grained motions

asked 2018-11-27 16:45:15 -0600

kdx2 gravatar image

updated 2018-11-27 16:46:46 -0600

Hello community, I am trying to analyze small facial expressions of around 0.5s duration around the mouth, the nose and the eyes. Could you please tell me which optical flow algorithm suits my task best?

edit retag flag offensive close merge delete



why do you think, optflow is relevant here ?

berak gravatar imageberak ( 2018-11-27 19:17:32 -0600 )edit

@berak I read a paper, where they are using CNN + RNN in which they feed the optical flow matrices to boost the prediction accuracy and I was just trying to simulate it better. If you have any other idea, or if you could expand on what you mean by xy problem, I am eager to read it carefully.

kdx2 gravatar imagekdx2 ( 2018-11-28 04:08:13 -0600 )edit

oh, i might have been simply wrong above ;)

berak gravatar imageberak ( 2018-11-28 04:22:49 -0600 )edit

@berak, no worries. But still, for the task I am trying to solve, which optical flow should I use best and shall I downsample myimages or that's gonna do no good? I am in love with the speed of DIS but there is a horrific amount of noise with my current settings and I don't know how to change them to make it work. TVL1 seems the best but is terribly slow in Python. Any suggestions will be welcomed. :D

kdx2 gravatar imagekdx2 ( 2018-11-28 04:50:59 -0600 )edit

do you have a link to the paper ? (just curious)

berak gravatar imageberak ( 2018-11-28 05:24:35 -0600 )edit

@berakhttps://arxiv.org/abs/1411.4389 , there you go. Btw, I started using the T-API, now is around 3-4 times faster. TVL really looks like the best option: ScalesNumber(2), ScaleStep(0.5), Tau(0.3), WarpingsNumber(2), InnerIterations(5). When the frame is idle sometimes a lot of noise starts being generated. Can you hint me how can remove/lessen it?

kdx2 gravatar imagekdx2 ( 2018-11-29 04:34:33 -0600 )edit