What are the Limits to Time Series Based Recognition of Semantic Concepts?

Most concept recognition in visual multimedia is based on relatively simple concepts, things which are present in the image or video. These usually correspond to objects which can be identified in images or individual frames. Yet there is also a need to recognise semantic concepts which have a temporal aspect corresponding to activities or complex events. These require some form of time series for recognition and also require some individual concepts to be detected so as to utilise their time-varying features, such as co-occurrence and re-occurrence patterns. While results are reported in the literature of using concept detections which are relatively specific and static, there are research questions which remain unanswered. What concept detection accuracies are satisfactory for time series recognition? Can recognition methods perform equally well across various concept detection performances? What affecting factors need to be taken into account when building concept-based high-level event/activity recognitions? In this paper, we conducted experiments to investigate these questions. Results show that though improving concept detection accuracies can enhance the recognition of time series based concepts, they do not need to be very accurate in order to characterize the dynamic evolution of time series if appropriate methods are used. Experimental results also point out the importance of concept selection for time series recognition, which is usually ignored in the current literature.