Budget-Aware Activity Detection with A Recurrent Policy Network
share
In this paper, we address the challenging problem of effi- cient temporal activity detection in untrimmed long videos. While most recent work has focused and advanced the de- tection accuracy, the inference time can take seconds to minutes in processing one video, which is computationally prohibitive for many applications with tight runtime con- straints. This motivates our proposed budget-aware frame- work, which learns to perform activity detection by intel- ligently selecting a small subset of frames according to a specified time or computational budget. We formulate this problem as a Markov decision process, and adopt a recurrent network to model a policy for the frame selec- tion. To train the network, we employ a recurrent policy gradient approach to approximate the gradient of the non- decomposable and non-differentiable objective defined in our problem. In the extensive experiments on two bench- mark datasets, we achieve competitive detection accuracy, and more importantly, our approach is able to substantially reduce computational time and detect multiple activities in only 348ms for each untrimmed long video of THUMOS14 and ActivityNet.
READ FULL TEXT