Using Principal Component Analysis (PCA) for feature reduction (HOG-PCA)

Question

Using Histogram of Ordered Gradients (HoG) I have computed features of 15 sample images. The feature vectors generated by these samples are very large (i.e. take up a lot of memory).

To reduce these feature vectors, i am using Principal Component Analysis (PCA). Here is the OpenCV code I am using:

PCA pca(imageT, Mat(), CV_PCA_DATA_AS_ROW, 300);
pca.project(imageT, imageT1);

In imageT Matrix, no.of row =no. of sample no. of coloumns = no. of features Suppose for 15 images
no.of row of imageT is 15 and no. of coloumn is 57400

I want 300 features after applying PCA; it gives me less than 15 features. I need help.

See also

• OpenCV Documentation: PCA
• Wikipedia Histogram of Ordered Gradients
• Wikipedia: Principle Component Analysis
• Selection of Histograms of Oriented GradientsFeatures for Pedestrian Detection
• Kernel PCA of HOG features for posture detection

Answer 1

You need to use more images for training. PCA captures variances that are orthogonal to each other. More data gives it the ability to find more basis vectors.