Computer Vision News - April 2016

COMPUTER VISION NEWS 17 An example of the flexible ordered graph , where the most recent action is take milk box, marked in red, and the predicted action is pour milk into mug. The possible paths from the most recent action to the predicted one are marked in green and yellow. Research Each recorded video is a sequence of actions and has an action and a boundary (where it start and ends). The videos are divided into fixed parts that represent action progress level. The researchers use discriminant model to train classifier to predict this progress level as action might have different time duration. The observation comes from individual progress level classifier for each action category – this unique model allows the coupling the recognition, segmentation and prediction at the same time. During inference a new sequence video is divided into the same fix length as used in training. Features are extracted to each part, and the action for each part is predicted by the model. The Flexible Ordered Graph To determine whether there is a missing action or not there is a need to extract possible action ordering, for this purpose the researchers used Flexible Ordered Graph (FOG). The FOG allows identifying a missing action by calculating the costs of the possible paths between the recently completed and the predicted action. Segmentation and Prediction Model To extract the underlying sequences of action the video the researchers uses the HMM graphical model. The states in the graph are action progress level and they represented the action category and if the action is in it beginning (B A ), middle (M A ) or end state (E A ).