BING3D: Fast Spatio-Temporal Proposals for Action Localization

The goal of this work is realistic action localization in video with the aid of spatio-temporal proposals. Current proposal generation methods are computationally demanding and are not practical for large-scale datasets. The main contribution of this work is a novel and fast alternative. Our method uses spatio-temporal gradient computations, a generalization of BING to the temporal domainleading to
BING3D. The method is orders of magnitude faster than current methods and performs on par or above the localization accuracy of current proposals on the UCF sports and MSR-II datasets. Furthermore, due to our efficiency, we are the first to report action localization results on the large and challenging UCF 101 dataset. Another contribution of this work is our Apenheul case study, where we created and tested our proposals performance on a novel and challenging dataset. The Apenheul dataset is large-scale, as it contains full high definition videos, featuring gorillas in a natural environment, with uncontrolled background, lighting conditions and quality.