[Bf-committers] Revised GSoC proposal: Realtime motion capture input and motion recording

[Lars Krueger]

> I have modified my proposal a lot according to the comments and advices of
> the developers on the mailing list.
> Here is link of the modified proposal.
> 
> Realtime motion capture input and motion recording for Blender - Test case
> is using the Kinect. <http://wiki.blender.org/index.php/User:Shuvro>
> 
> Thanks to all of you for your kind help and advice.

Instead of implementing a classifier yourself, make use of readily available libs, e.g. libsvm, SNNS, or libANN

Regarding k-nearest-neighbour (kNN) vs. neural network: Which one is more robust depends mostly on the distribution of you input data. As a rule of thumb: if you have several randomly distributed clusters, kNN is better. If you have connected "peninsulas", neural networks such as the MLP are better suited.

Without reading the paper, I would assume the authors used kNN due to its performance for this problem. I could imagine that many people put their foot down in only a few way, while there are quite a few ways to move a foot while not putting it down. Hence, you have a few clusters in the "foot down" class, and basically random samples for the other class(es). 

Before jumping to a different classifier, why not simply re-implement the paper first? A kNN with few classes is pretty hard to beat regarding speed, mostly only by a polynomial classifier.

Also: Backprop is a training method for neural networks, which can be applied to a number of NNs. It is not a neural network itself as your wiki entry suggests. You might want to correct that for the application. I guess the guys at Google are well-versed regarding classification.

-- 
Dr. Lars Krueger


Empfehlen Sie GMX DSL Ihren Freunden und Bekannten und wir
belohnen Sie mit bis zu 50,- Euro! https://freundschaftswerbung.gmx.de