The difference between the 3D Inception Module and the 2D.

The Inception_v1 have 3x3 and 5x5 convolution layer and the bn_inception have 3x3 and two 3x3 conv layer in the middle branch in each inception module. But it the 3D inception module, you only save 2 3x3 conv layer, one for each branch. Can you tell me why you do this and when you transfer 2D bn-inception parameters to 3D model, do you just ignore the second 3x3 conv layer in the second branch?
Thanks!