目标跟踪中基于深度可分离卷积的剪枝方法

Pruning Based on Separable Convolutions for Object Tracking

为了减少跟踪网络中存在的参数量和计算量大的问题,提出了基于深度可分离卷积的剪枝方法。深度可分离卷积将跟踪网络中的传统卷积层分解为逐点卷积和逐层卷积两部分。在逐点卷积中,通过逐点卷积层中权重的大小来评估输入特征图通道在线性组合中的重要程度,将较小的权重及其关联的特征通道裁减掉。在逐层卷积中,通过K-L散度来衡量逐层卷积中滤波器的相似性,将相似的滤波器裁剪掉,减少冗余。通过上述方法进行多轮迭代剪枝,从而减少跟踪网络的参数量和计算量。在VOT数据集上的实验结果表明,在精度没有下降的前提下,剪枝后网络的参数量下降了22.54%,计算量下降了17.8%。在NVIDIA TX2设备上的实验结果表明,剪枝后网络的跟踪速度在CPU上提升了14.95%,在GPU上提升了13.07%。

A pruning method based on separable convolution is proposed to reduce parameters and calculation amount in the tracking network.Separable convolution decomposes the traditional convolutional layer into two parts,i.e.pointwise convolution and depthwise convolution in the tracking network.In pointwise convolution,the importance of the input feature map channel in the linear combination is evaluated by the weights in pointwise convolutions,so that the smaller weights and their associated feature channel ought to be pruned.In depthwise convolutions,the K-L divergence is adopted to measure the similarity of filters,and the similar filters are then pruned to reduce redundancy.With the improvements above,iterative pruning is performed.The parameters in the tracker are reduced by 22.54%,and the computation is reduced by 17.8%.On the NVIDIA TX2,the tracking rate increases by 14.95%for CPU and 13.07%for GPU without significant degradation of tracking performance in VOT benchmark.