What method is used from 2D convolution expansion to 3D convolution？

[luoji-ai]

As you said on your twentybn community blog, "We took inspiration from Carreira and Zisserman (2018) and temporarily inflated 8 2D convolutions, effectively converting them to 3D convolutions", from 2D convolution to Is the three-dimensional convolution performed using the temporal shift module (TSM)?