Learning Better Image Representations Using ‘Flobject Analysis’: Supplementary Material

There are four major stages in the flobject pipeline: preprocessing, flobject analysis, creating image descriptors, and classification. Before any analysis is done, video frame pairs and static images are first preprocessed and reduced to a suitable representation during the preprocessing stage. Next the unsupervised flobject analysis stage takes as input a collection of video frame pairs and the number of topics to learn, and returns as output a distribution over codewords for each topic. After analysis, the learnt per-topic distributions over codewords are used to create an image descriptor for each static image that is useful for classification tasks. Finally, in the last stage, the computed image descriptors for a collection of labelled training images and unlabelled test images are fed into a standard classifier to predict the object class for the test images.