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.