This is a quick and easy implementation of a head pose estimator based on geometric model fitting via 2D-3D points correspondence. Results aren't that good, but for a 5 min work and practically no theoretical knowledge - they're pretty nice.
See the blogpost for more details and the code
http://www.morethantechnical.com/2010/03/19/quick-and-easy-head-pose-estimati...
Angelina's pictures are from LFW
http://vis-www.cs.umass.edu/lfw/person/Angelina_Jolie.html
If you consider the narrow range of your samples - having pitch yaw and roll pretty much constrained within 10 degrees - your error is pretty big.
ivionday 1 month ago
can you explain the 3x3 rotation matrix?
e.g. how do 9 discrete values translate into 3 values- pitch roll and yaw?
kosomonova 9 months ago
Looks like the results are just ok, especially in yaw (looking left to right) estimation
ivionday 1 year ago