Experimental tasks

List of tasks (8/8/2012)
  

• Verify the correctness of implementation on all processes of existing work

• Done usingKTH and Weizmanns action dataset
• Benchmark against VLFeat's BOW using Laptev's HOG-HOF descriptor
• List of different processes involved:
• STIP detection
• Number of interest points
• BOW formation
• Vector quantization
• Histogram binning
• Building of classifier
• ST Correlation & Haralick texture computation

• Computational complexity is to be calculated
• MMI Clustering
• Evaluate performance of K-means against MMI clustering
• Try to use MMI clustering as the primary clustering process

• Verify behaviour of classifier against unknown action datasets as well as action classes
• Inter dataset classification
• Behaviour against unknown action class, abilitiy to distinguish unknown action classes

•  Research on other texture measurement methods which are more recent in comparison to Haralick