Classifier with a lot of false positive

asked 2017-01-19 08:01:34 -0500

Hello, I'm depressed ... It's been a week since my training classifier turns and it's the drama ... I try to create a classifier to detect magpies ...

opencv_traincascade -data classifier -vec positives.vec -bg negatives.txt -numPos 1000 -numNeg 600 -numStages 10 -w 48 -h 48

I had a lot of trouble for the samples because the magpie is black and white, when creating samples there is a part of the magpie that disappears.

I was finally able to create samples using the script annotations, and then my samples.

I have 2000 positive images and 600 negatives. After a week of training, I test on an image and it works but I have detections more.

I have to mount the "minNeighbors" parameter thoroughly to make them disappear. Do you have any ideas to improve my classify?


edit retag flag offensive close merge delete



do you really have 2000 positive images ? or did you generate synthetic ones from a single image in create_samples (bad idea) ?

berak gravatar imageberak ( 2017-01-19 08:13:03 -0500 )edit

i use a lot of images (~200) with the annotation script who gives me the positions of the object (F... Magpies), and after, i put them in samples with create_samples who give me 1000 positives images

kabdriver gravatar imagekabdriver ( 2017-01-19 10:35:38 -0500 )edit

oooh, ok, i really did not expect that ..

berak gravatar imageberak ( 2017-01-19 10:45:44 -0500 )edit

Like he said, bad choice, read OpenCV 3 Blueprints, chapter 5 to know why :)

StevenPuttemans gravatar imageStevenPuttemans ( 2017-01-19 10:49:44 -0500 )edit

again, "birds-in-the-wild" are terribly hard.

maybe you can restrict it to a "profile" view, beak on the left ? (and flip the image later, to detect the "opposite" pose)

berak gravatar imageberak ( 2017-01-19 10:54:59 -0500 )edit