What images to use for negative training data for SVM?

asked 2017-03-02 21:04:39 -0600

I'm currently using HOG (Histogram of Oriented Gradients) to generate feature vectors for each of my training images, and then using this data to train an SVM Classifier. Since the SVM classifies the incoming data into one of 2 classes, I understand I need to provide training images for both classes.

I hope to use this to detect pedestrians, and I have a set of a few hundred 128x64 images of different pedestrians. However, I'm not sure what images to use for the other class, i.e. the non-positive one. Should I just use images of different ambient scenes with no pedestrians in them?

edit retag flag offensive close merge delete

add a comment

1 answer

Sort by » oldest newest most voted

answered 2017-03-03 01:06:39 -0600

berak
32993 ●7 ●81 ●312

Should I just use images of different ambient scenes with no pedestrians in them?

yes, that's a start. maybe you can even crop portions of the original images, that do not have any humans in it.

edit flag offensive delete link

Comments

Or you simply take a look at available datasets, like INRIA, CAVIAR, ... that have pedestrian datasets and use their negative training data. Do not reinvent the wheel I would say :D