http://people.csail.mit.edu/celiu/OpticalFlow/
Mr. Liu uses an algorithm that combines minimizing local energy functions from Lucas-Kanade and a global energy function from Horn-Schnuck to get dense flow that is resilient to noise. I tried a similar algorithm earlier on, but I suppose they were just of inferior quality because this algorithm is doing extremely well.
Left Image
Right Image Warped
My one problem moving forward is still speed. It will take a few minutes to fully register a set of images, which is OK, but a little annoying with very large databases. If I can constrain the objective function in Liu's algorithm with calibration data, I can get a major performance boost. Liu is using some pretty heavy mathematics that I'll have to spend some time learning if I'm going to be able to make changes confidently.
No comments:
Post a Comment