A Non Local Multi-Fiber Network For Action Anticipation In Videos

Nour El-Din El-Madany, Lei Gao, Ling Guan

Abstract: In this paper, we investigate the fascinating capacity of 3D Deep ConvNets in action anticipation. Specifically, a Non Local Multi-Fiber Network (NL-MFN) model is proposed. The proposed model is compared against several 3D Deep ConvNets for the task of early action recognition or partially observed frames on the available UCF101 dataset. Additionally, we present the results on the EPIC kitchens challenge dataset used in action anticipation task at the CVPR conference. The experimental results demonstrated the effectiveness of the proposed NL-MFN in action anticipation.