Pose estimation

[bnascimento]

How can i obtain the pose estimation for a given face?
I understand that has to be with the best mixture, maybe?
Would aprecciate a little help, ty

[bnascimento]

i've now understand that component_ member of a candidate indicates the "yaw" given by the view point model.
How about the head position "pitch" value given eigen vectors 1 and 2?
How can i calculate those?

[bnascimento]

Still can't find any way to get pose estimation directly...
is there a way to get pitch, yaw, rolll from all the candidates?

[vrabaud]

Just curious: is there a paper dealing with that ?
On Mar 19, 2013 4:16 PM, "bnascimento" [email protected] wrote:

Still can't find any way to get pose estimation directly...
is there a way to get pitch, yaw, rolll from all the candidates?

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/8#issuecomment-15120472
.

[bnascimento]

This work is based on http://www.ics.uci.edu/~xzhu/face/
whose title is "Face Detection, Pose Estimation and Landmark Localization in the Wild". Face detection and ladmark location is possible, but pose estimation is not direct.
My initial guess is that pose estimation would be possible through the candidates winner model/mixture. See videos on the project website. So far I was able to get yaw values by obtaining the winner model, but I cannot get the head position "pitch" value given eigen vectors 1 and 2 because i cannot find them...

[vrabaud]

So you need a new function in the API ? What would the new signature be ?

[bnascimento]

Yeah a new function that returned the euler angles for head pose (yaw, pitch and roll) would be awesome.
Something like: EstimatedHeadPose(double & yaw, double & pitch, double & roll).
Roll angle can be given directly by the two eyes angle with an horizontal line.

[vrabaud]

This is very specific to faces. @hbristow is busy these days but as far as I remember the code is very generic compared to your implementation. I was more thinking of an API function that would expose some more info (like eigen values or something) and on which you could build your work.
BTW, just realized I had closed that issue but I probably meant to close another one. Sorry, re-opening.

[hbristow]

This is indeed a very specific request for what was intended to be a general recognition architecture.

I agree that some form of pose estimation would be good, and while I think it would be best if you write the few lines of code to extract pose specific to faces, I can certainly help out by providing a more verbose output from the winning candidate. What kind of information are you interested in?

H

On 23/03/2013, at 5:19 AM, Vincent Rabaud [email protected] wrote:

This is very specific to faces. @hbristow is busy these days but as far as I remember the code is very generic compared to your implementation. I was more thinking of an API function that would expose some more info (like eigen values or something) and on which you could build your work.
BTW, just realized I had closed that issue but I probably meant to close another one. Sorry, re-opening.

—
Reply to this email directly or view it on GitHub.

[bnascimento]

Hi, yes I agree about this beeing a very speficic request for face detection. Maybe a more generic approach would be better.
My main interest would be to expose more information about the winning model, e.g. which mixture won / has more candidates points / the score of this candidate points, etc...
My main consern would be the eigen value 1 and 2 described in the videos on the project website which represent the head pitch angle. I cannot find them under this implementation. For yaw angle values im using the winning model...
Thank you for helping