Investigating time-sensitive topic model approaches for action recognition

AbstractIn this paper, we present several attempts of using topic models for ac-tion recognition in videos. We show that time-sensitive topic models helprecognizing actions when little training data is available. We also exhibitsome limitations of these models when dealing with complex videos. Newapplications of these models in semi-supervised settings and the use of in-herently discrimant models such as the MedLDA one are also considered. 1 Introduction Action recognition is an important eld of video processing. Its applications cov-ers, among others, automatic annotation of videos, improved human-computerinteraction and guidance in monitoring public spaces. Most state-of-the-arttechniques for action recognition video documents rely on Bag-of-Word (BoW)representations. The latter are built from quantized spatio-temporal descriptorscollected over long video segments [10, 14, 12, 13]. Such methods, however, donot encode the time information, although actions are characterized by strongtemporal components. To address this issue and enhance action recognition per-formance, we investigate the use novel principled probabilistic methods (calledtopic models) for capturing the temporal relationships between characteristicsub-units of a given action. In a previous paper [17], we showed these modelscould help action recognition when little training information is available. Inthe following, we expose more experiments to better spot strong points andweaknesses of these models when used for action recognition.