train a classifier based on the huge face database presented by: jie chen jie chen, ruiping wang,...
TRANSCRIPT
Train a Classifier Based on the Huge Face
Database
Presented by: Jie Chen
Jie Chen, Ruiping Wang, Shengye Yan, Shiguang Shan, Xilin Chen, Wen Gao
Motivation Data collection is tedious but essential for
learning based algorithms In Viola CVPR 2001, bootstrap for negative; Ours: Resampling the positive set, besides the
bootstrap for negative.
Why? Collected face samples randomly;
Result in the bias of the trained detector.
How? Fill in the face example space by GA; Subsample it by manifold; Mend by SVM.
The collected face and nonface
set
nonface set
face set
nonface set
face set
Resulting distribution.
Contribution of this Paper Subsample a small but efficient and
representative subset based on the manifold: Discuss the effects of outliers; The performance is instable to train a detector based
on the random subsampling. However, a detector trained on the subsampled face set by manifold is not only stable but also performance improved;
When we prepare the training set, we should collect more samples along those dimensionalities with larger variances to get a nearly uniformed distribution in the manifold, for example, left-right pose of faces more than up-down pose.
A typical manifold – Swiss Roll(B. J. Tenenbaum, V. Silva, and J. Langford )
Manifold
from http://www.cs.toronto.edu/~roweis/lle/
Face Sample Manifold
An individual with varying pose and expression
from http://www.cs.toronto.edu/~roweis/lle/
Too dense!Too sparse!
Dimensionalities of Isomap The residual variance of Isomap
embedding on the 698 face database
lighting direction
up-down pose
left-right pose
Dimensionalities Each coordinate axis of the embedding correlates
highly with one degree of freedom underlying the original data:
left-right pose corresponding to the first degree of freedom;
up-down pose corresponding to the second one ; lighting direction to the third one.
That is to say the scatter of face images in left-right pose is the biggest while the scatter in lighting is the smallest among these three factors.
We conclude that, in order to select representative example set, we should pay more attention to the left-right pose variations than the up-down pose.
Subsampling by manifold
1
3
2
(a) illustration of subsampling based on the estimated geodesic distance; (b) manifold of 698 faces; (c) subsampled results.
(a) (b) (c)
Experiments:Subsampling by manifold training set -- 6,977 images (2,429 faces
and 4,548 non-faces) testing set -- 24,045 images (472 faces
and 23,573 non-faces). All of these images are grayscale and
they are available on the CBCL webpage. let K=6 for the manifold learning. Trained on the AdaBoost based classifier
Subsampling based on manifold
Some possible reasons: Examples subsampled based on the manifold distribute r
easonable in the example space and have no example congregating compared with the whole set;
Outliers in the whole set deteriorate its performance
Large scale of database The face-image database consists of 100,000
faces (collected form web, video and digital camera);
Randomly rotate , translate and scale; After these preprocessing, we get 1,200,000 face
images which constitute the whole set; The first group is composed of 15,000 face
images which are subsampled by the manifold (ISO15000) ;
The second or third group is also composed of 15,000 face images which are random subsampling (Rand1-15000 and Rand2-15000).
Test on MIT+CMU set Sampled training set by the manifold and
the random subsampled set Trained on the AdaBoost based classifier
Conclusion Present a manifold-based method to subsample.
Compared with the detector by random subsampling, the detector trained by manifold is more stable and achieve better performance.
Improved performance results from: Reasonable-distributed examples, subsampled based on
manifold, No outliers, discarded during the manifold learning