Bag of Words model for gesture recognition

Article: Measuring and Reducing Observational Latency when Recognizing Actions http://www.cs.ucf.edu/~smasood/publications/ICCV2011_LatencyReduction.pdf

Dataset: http://www.cs.ucf.edu/~smasood/datasets/UCFKinect.zip

[Part 6.1.1]: Bag of Words Model The bag of words is computed using the same distances described in Section 5. The expansion to binary features was not used because the raw distance values performed best on this classifier. The BoW representation is created by quantizing the feature representation of each frame to one of 1000 clusters. The clusters were chosen randomly from the dataset. This had similar performance to using the k-means algorithm to find the centers, but was significantly faster. Each video is represented by a histogram describing the frequency of each cluster center. Histograms are normalized to avoid bias based on the length of the video. The classifier is implemented using an SVM based on the histogram intersection kernel.

LibSVM with Intersection + Chi Square Kernels: http://dovgalecs.com/blog/libsvm-3-0-matlab-with-intersection-kernel/

Experiment source: experiment.m