Robust Compression Technique for YOLOv3 on Real-Time Vehicle Detection

For vehicle detection, YOLOv3 has shown promising accuracy. Since the number of parameters in this network can be more than ten million parameters, it cannot be fit into a commodity camera. In this paper, we propose a compression mechanism designed specifically for YOLOv 3's network by removing unnecessary filters. Since YOLOv3 composes of two network components: backbone and pyramid networks, we propose a robust pruning mechanism to prune filters of each network separately. This can help to avoid over-pruning the network in some part of the model making our model more robust. There are two main pruning criteria investigated: Average Percentage of Zero (APoZ) and Sum Magnitude Weight. The experiment was conducted on UA-DETRAC. The results show that our compression mechanism with APoZ criterion can reduce more than 90% of the network size, while the accuracy is even higher than the full model for about 2%.