An effective hybrid pruning architecture of dynamic convolution for surveillance videos

Abstract

The large-scale surveillance videos analysis becomes important as the development of the intelligent city; however, the heavy computational resources necessary for the state-of-the-art deep learning model makes real-time processing hard to be implemented. As the characteristic of high scene similarity generally existing in surveillance videos, we propose an effective compression architecture called dynamic convolution, which can reuse the previous feature maps to reduce the calculation amount; and combine with filter pruning to further speed up the performance. In this paper, we tested the presented method on 45 surveillance videos with various scenes. The experimental results show that the hybrid pruning architecture can reduce up to 80.4% of FLOPs while preserving the precision within 1.34% mAP; furthermore, the method can improve the processing speed up to 2.8 times compared to the traditional Single Shot MultiBox Detection.