Wednesday, August 22nd, 2012, 276 days ago

Collection of own hand gesture samples – the call for pictures, training method, tool and results

The call for hand gesture samples

Quite long time ago a call for hand gesture samples was made in here, unfortunately this call received few responses making contributions to the samples. Guess if I post free databases of hand gestures for high quality HAAR training then the situation would be quite different – kidding.

The quality of samples matters!!!

Of course high quality samples can improve the training results, but the real situation is that it is not easy to take thousands of hand gesture photos in a short time and these photos need to be in a series and include target detection features. Plus cropping these features out of photos manually, this is overloaded. So that is why we need some hand gesture databases to do the job. In my project I had a chance to use Cambridge’s hand gesture database, these samples were in 80×60 size and had most hand gestures. However the backgrounds were black and the features were not very well cropped, so I have to give up, and instead, starting to collect my own samples. I wish someday I could publish these as another hand gesture database sharing with the public.

Based on my experiences of training so far, the samples from webcam used to have good enough quality for the training. Normal webcams can produce 320×240 pictures, and nowadays 640×480 is quite common. With the webcam, then with a short programme using openCV to grab webcam streams and save these as specifically sized samples, the sample collection work is done. An example of using openCv to collect sample pictures can go here Writing a simple hand gesture picture sampler for classifier training. Also, in Naotoshi’s tutorial, another tool was provided to speed the manual cropping of sample pictures, the link is here image cropper

Negative samples ALSO matter

Negative samples are those pictures without target features, so theoretically any background pictures off the hands can be used as negative samples. Therefore this is easier to collect. What I intend to mention here is, I made a stupid mistake in my project and that led to very bad detection results, no matter how much positive samples were collected and cropped. I resized the background pictures to 40×30 size, and the positive samples were with 20×20 size. Based on that, the detection results generated unbelievably high false detection rate, as I mentioned in a previous post here A failed example of hand gesture recognition using openCV haartraining classifiers.

And after I found out and changed these negative samples to normal sizes like 320×240, bingo~ I got quite good improvements, and I could see the differences of detection accuracy each time I added positive and negative samples for training.

QR code for this post, SCAN ME