|
This paper introduced a new framework, named Motion-enhanced Semi-3D (MeS3D), for surveillance video classification. Considering there was no public surveillance video classification benchmark for us to evaluate the performance of models, we made SV Dataset that is a small dataset consisting of 1979 trimmed surveillance videos. With two branches extracting motion information separately from videos, our MeS3D achieved higher accuracy on SV Dataset than state-of-the-art, demonstrating that the proposed MeS3D is a novel feature learning framework for surveillance video classification. |
|
Keywords:Deep learning, Motion-enhanced Semi-3D, surveillance video classification, SV Dataset |
|