Have Geek Brain

Machine learning - learning step

When you use Haar-like functions on your training data for the Adaboost algorithm, how do you create your datasets? Do you literally have to find thousands of positive and negative samples? There must be a more efficient way to do this ...

I am trying to parse images in matlab (not faces) and am relatively new to image processing.

+2


a source to share


4 answers


Yes, you need a lot of positive and negative learning patterns. This is especially true for Adaboost, which works by resampling the workout set. How many samples are enough is hard to say. But in general, the more the better, because it increases the chances of your training set being representative.



Also, it seems to me that your quest for efficiency is misplaced. Training is conducted in advance, presumably offline. The efficiency of classification of unknowns after training is done, which is what people usually worry about.

+4


a source


Surely more data, more information, better results. You should include more details. However, one thing you may need help with is the ratio of the positive set to the negative set. For logistic regression the ratio should not exceed 1: 5, for adaboost I'm not sure about the result, but it will certainly change with the ratio (I've tried before).



0


a source


Yes, we need a lot of positive and negative samples for training, but collecting this data is very tedious. But you can do it easily by taking videos instead of images and using ffmpeg to convert those videos to images. This will make the tutorial part easier.

0


a source


The only reason to have the same positive and negative samples is to avoid bias. Sometimes you can get high accuracy, but not completely categorize one category. For evaluating such methods, accuracy / recall is more useful than accuracy.

0


a source






More articles:


All Articles