Need to detect Tiger faces in its natural habitat. Using inbuilt Haar-training (opencv_haartraining) feature to train the system but the performance is low. How to increase the hitrate while reducing the false alarm rate?
More training and especially a better training of the background, by using more negatives is the first I should do. Secondly, try other classifiers, like SVM, random trees, ...
For more classifiers, check this link:
As the number of stages increases the hitrate as well as the false alarm rate increases. How to avoid this? Usage of 2200 negatives against 500 positives have been made, is that sufficient?? The OpenCV Haar training has an inbuilt cascade of weak classifiers with the Adaboost, how can we change this classifier?
Basically a cascade classifier is made with boosting, like adaBoost. If you want another classifier, you just need to build a new system yourself. Steps to follow are for example :
Calculate feature points of the image (SIFT or SURF)
Create descriptors around those feature points
Store desccriptors in a codebook
Create a SM classifier with this data
Basically you need to google these items and read, so that you understand what happens!
Asked: Apr 6 '13
Seen: 574 times
Last updated: Apr 06 '13
