Hello everyone, I'm currently working on my master thesis in compressed sensing for video. Part of my project might be to learn a dictionary from a large collection of videos. The typical algorithm to use for this task is called K-SVD. K-SVD is very similar to K-Means but there are two key differences: - A data point is assigned to several 'clusters' (not one) - Instead of finding the cluster by taking the mean of all data points the cluster is found by finding the first principal component using an SVD (hence the name). The original reference for the algorithm is "K-SVD: An Algorithm for Designing Overcomplete Dictionaries for Sparse Representation" from M. Aharon.
It seems no one has implemented this algorithm yet on Mahout. Since I have no experience with Mahout/Hadoop I wondered if it would be difficult to implement this based on the available implementation of K-Means? Is anybody interested in this and could point me towards the right direction for an implementation? Best, Thomas
