OpenCV Bad trained Hog SVM

Asked Jan 06 '15 at 18:33

Active Jan 06 '15 at 18:33

Viewed 429 times

I found this repository for training a SVM with HOG features. I changed the code so the hog window size is not 64x128 but 64x64 instead as I don't want to detect people. My goal is to detect a car steering wheel. After training I have:

True Positives: 303
True Negatives: 461
False Positives: 6
False Negatives: 2

My problem is that while testing with my webcam, I got false positives. It detects every circular shape and also some other shapes which doesn't look like a steering wheel at all. The circular shapes are well detected (as I mean that while moving them and rotating them they are still detected) while the other false detected are detected only in a specific position.

Do you have any idea how can I train the SVM correctly?

asked Jan 06 '15 at 18:33

dephinera

3,703
11
41
75

Maybe circular shape is the only thing thats in common for most of the steering wheels? How many samples did you use? – Micka Jan 06 '15 at 18:48
I use 305 positive samples and 379 negative – dephinera Jan 06 '15 at 18:50
You can try to retrain and add the false positives (called "hard samples") to the negative samples! But keep in mind not to use those samples for testing afterwards. – Micka Jan 06 '15 at 18:51
I mean to add false positives of your webcam testing to the retraining – Micka Jan 06 '15 at 19:41
1

I get your idea. Thank you for your help. I write again when I try it :) – dephinera Jan 06 '15 at 20:06

OpenCV Bad trained Hog SVM

0 Answers0