LPM for Action Recognition in Temporally Untrimmed Videos

In this notebook paper, we evaluate GBH and MBH descriptors for action recognition in temporally untrimmed videos. Our system is based on the recent improvement of local part model with gradient boundary descriptor [5]. We extract both local GBH and MBH descriptors and represent them with Fisher vector. We use LPM to include local structure information. We apply a slide window approach to extract short clips from temporally untrimmed video, and using a linear SVM to classify each short clip. We simply label the untrimmed video using the clip with the maximal classification score.