MouseGestures - Neural Network Project ETSII ULL 01

[SlykeThePhoxenix]

That's actually really cool.

[maccat]

very amazing :O

[darkdragon7099]

Amazing, I wanna get started on this subject. Any tips?

[botrules]

For recognizing gestures I have used the Neural Network approach. There is sure other ways to do it but I found NN very interesting .

If you take the path of NN you have to learn about it first to catch the most important concepts. I thinks little teory its important before trying something. After that you are ready to code something you want be resolved by a NN.

Hope being helpful.

[biorythme232]

Nice work. What is the type of neuronal network ? Perceptron ? how many layers and neurons ?

And what's the input data, the pixel positions ?

[botrules]

I have used a Multi Layer Perceptron.

The structure is 50 float inputs, 25 middle layer neurons and a dinamic size output depending the number of different gestures.

The input data are not the pixel positions. Its more or less the difference of one pixel respect to the next one. Math concept of derivative.

You can take a look at the source code, linked on the description.

[biorythme232]

Your choice about the data to use is a very interesting and efficient approach. Thanks for the link !

[AndrianGo]

I don't get it. What does the derivative give you? why is that better than direct input?

[botrules]

Direct input is not good because Neural Network would recognice only the same movement. I mean, the same movement size in the same 2D coordinates. Thats not really what we mean with gesture.

Working with the derivate gives you more or less the gesture data from the movement. Size and 2D coordinates are not important.

[AndrianGo]

Are you making vectors from each consecutive pairs of pixels and then feed them to the NN? Is that how you think of it as derivative?

[botrules]

The code pseudocode is:

gestureDataX[i] = moveData[i].getX() - moveData[i - 1].getX()

gestureDataY[i] = moveData[i].getY() - moveData[i - 1].getY()

[AndrianGo]

Makes sense now :)

Thank you

[d4mr0d]

muy buen trabajo, enhorabuena!