Learning Spatial and Temporal Extents of Human Actions for Action Detection

For the problem of action detection, most existing methods require that relevant portions of the action of interest in training videos have been manually annotated with bounding boxes. Some recent works tried to avoid tedious manual annotation, and proposed to automatically identify the relevant portions in training videos. However, these methods only concerned the identification in either spatial or temporal domain, and may get irrelevant contents from another domain. These irrelevant contents are usually undesirable in the training phase, which will lead to a degradation of the detection performance. This paper advances prior work by proposing a joint learning framework to simultaneously identify the spatial and temporal extents of the action of interest in training videos. To get pixel-level localization results, our method uses dense trajectories extracted from videos as local features to represent actions. We first present a trajectory split-and-merge algorithm to segment a video into the background and several separated foreground moving objects. In this algorithm, the inherent temporal smoothness of human actions is exploited to facilitate segmentation. Then, with the latent SVM framework on segmentation results, spatial and temporal extents of the action of interest are treated as latent variables that are inferred simultaneously with action recognition. Experiments on two challenging datasets show that action detection with our learned spatial and temporal extents is superior than state-of-the-art methods.