search

mz2 / MPGestures

Fork of https://github.com/fe9lix/DollarP_ObjC ported to OSX, with a supervised learning layer to improve gesture detection accuracy and backed by a crowdsourcable database of gestures
Other
21 stars 5 forks source link

Make resampling behaviour when translating configurable and evaluate options other than 32 #8

Closed mz2 closed 10 years ago

mz2 commented 10 years ago

The original DollarP code forces 32 point resampling -- this is not an arbitrary value as the Vatavu et al 2012 paper shows that there are little gains in their testing with going beyond 32 points and the runtime of the algorithm scales O(N^2.5) -- see http://faculty.washington.edu/wobbrock/pubs/icmi-12.pdf for more detail (basically, 5+ participants, 10+ training samples from each, with shapes resampled to 32 points). However, this needs systematic testing in a supervised setting because the performance effect of downsampling too much might only show itself in some symbol namespaces we make predictions in (more complex, curved shapes might need more samples).

mz2 commented 10 years ago

Refactored the point cloud creation API to include in all cases the resample count, scale normalisation and translation to a difference from the centroid option.

mz2 commented 10 years ago

The evaluation of different resampling settings was moved to a separate issue #9.