Frag HMM is a type of HMM taking the observation as the
grouping result of a low level over segmentation output. The link above
contains the matlab source code of fraghmm and example datasets. The code is based on the
HMM source from Prof. Kevin Murphy. His Original Source code
can be found here. The dataset is the generated groups from
the Two Handed Datasets.
The fraghmm has the ability to take all the
noisy segmentation considered. For example, in the experiment of the Two
Handed Datase. Although we have hundreds of noisy
groups generated, we can still achieve above 90% recognition rate. This
experiment can be reproduced using the above source code.
For example, to test the fraghmm with a
sum-sum approach, please run:
To test with a max-max approach, please run:
For more details about fraghmm. Please check
out the following paper:
R.D. Yang. and S. Sarkar, "Gesture
Recognition using HMM from fragmented observation ", IEEE
International Conference on Computer Vision and Pattern Recognition 2006.