Slim-SSD: An Efficient Fast Object Detection Method

Object detection networks require large storage space and high computational cost, it is difficult to deploy deep neural networks on embedded devices with limited memory and computing resources in actual object detection tasks. In order to solve these challenges, we propose an efficient object detection framework by pruning channels of the feature extraction layer of the network. First, we apply the L1 regularization to the channel scale factor in the BN layer to obtain the object detection network with sparse structure. Then we trim the channel with less information to get the object detection framework. Based on this method, we obtained Slim-SSD with less trainable parameters and test time. Our experiments on benchmarks show that, on the basis of approximate accuracy with the original network, we have reduced the number of model parameters by 3x, and reduced the testing time by 2x. Our compression method helps to deploy complex object detection networks on resource-constrained embedded platform.