《表1 不同卷积的计算量比较》

  提示：宽带有限、当前游客访问压缩模式

本系列图表出处文件名：随高清版一同展现

《基于深度时空卷积神经网络的点目标检测（英文）》

1. 获取 高清版本忘记账户？点击这里登录

1. 下载图表忘记账户？点击这里登录

The computation comparison of 3D convolution and factorized 3D convolution is listed as Table 1.The number of parameters of traditional 3D convolution with 10243D kernels is about 28.67K，and the requirement of computation is about 0.06GFlops.It can be approximated by 32 2D kernels and 32 1D kernels.As a result，only 9.38K parameters and 0.03GFlops are required by factorized 3D convolution.So，the factorization can reduce the computation burden by 50%.