Temporally enhanced image object proposals for online video object and action detections

摘要

Despite the recent advances of image object proposals (IOPs) and video object proposals (VOPs), it still remains a challenge to apply them to online video object/action detection. To address this problem, we propose a novel form of image object proposals, Temporally Enhanced Image Object Proposals (TE-IOPs), for online video object/action detection. The proposed TE-IOPs augment the existing IOPs at every frame by their temporal dynamics in the past few frames. We develop a dynamic programming scheme to efficiently search for such TE-IOPs in an online manner. Compared with existing VOPs that cannot run online, our TE-IOPs can be used for online detection. Compared with IOPs, our TE-IOPs bring rich temporal dynamics with minor computational cost. Experiments on benchmark datasets validate the superior performance of the proposed TE-IOPs over existing IOPs and VOPs, in terms of both the proposal re-ranking and the application of online action detection.