HW2 quirks

10 views
Skip to first unread message

Antonio Juarez

unread,
Oct 14, 2011, 1:53:59 PM10/14/11
to mlsp-fa...@googlegroups.com
Has anyone else had the problem where the simple dot product of the first 64X64 eigenface gets you better results than the complete boosting algorithm over the first 10 19X19 eigenfaces? So far I can only attribute this to the latter having such a low resolution, but I thought to ask, in case anyone else ran into the same issue.

Also, instead of up-sizing eigenfaces to scan different face sizes in the pictures, I've been down-sizing the pictures themselves. That should be fairly equivalent, even using the eigenface and error coefficients in the boosting algorithm, right?

Comments?

bhiksha raj

unread,
Oct 14, 2011, 7:58:00 PM10/14/11
to mlsp-fa...@googlegroups.com
The Eigenface weights by themselves aren't that great features.
Include the error too
(i.e. ||I - sum_k w_k eigenface_k||). This is a good indicator of
how good the Eigen faces are at explaining the image (which is a
pretty good predictor of whether its a face or not).

-Bhiksha

--
Bhiksha Raj
Associate Professor
Carnegie Mellon University
Pittsburgh, PA, USA
Tel: 412 268 9826

Reply all
Reply to author
Forward
0 new messages