I would like to count the number of fishes in a given image. All the fishes are from the same specie, so they are more or like the same shape.
This is an image example of what i'm going to see (not so blurry): http://i49.tinypic.com/w19uvn.jpg
Is it ok to train haar cascade or can I use another method (more accurate/faster) ?
If you want to detect any fish (not any specific specie), you can use binarization and blobs. OpenCV have a function that count the total number of blobs.
But if you need to detect only a specific specie of fish (in a picture that have other species), maybe haartraining is enough, but I don't know if it will get good results.
Agree. In the image attached intensity gradients are weak. That means that Haar will not be able to detect important features. It also performs bad if the object is rotated, but it is normal for fish. So, I agree that you should consider some custom algorithm, like blob counting.
You should probably forget about fishes and try general object counting approaches. Check people counting approaches: http://scholar.google.ru/scholar?&q=people+counting.
Thank you for your answers. I will try using blobs, indeed I only need to count fishes and they are the only objects that will appear in the images.
One last question. If some fishes are partially overlapped, is it still possible to get good results with blobs ?
Overlapping fishes is the difficult case. You can try erode blobs before counting to separate blobs from each other.
Asked: 2012-06-27 09:36:20 -0600
Seen: 4,641 times
Last updated: Jun 27 '12
