How does the 47τ in Table 1 of the paper come from?

[stephencheung16]

In figure 3 of the paper, each block of conv2_x, conv3_x, conv4_x, conv5_x and pool5 connects to 5 blocks of the previous layer.

I think T=3 in the two 1x1x3 conv accounts for the 5 connections from conv2_x, conv3_x, conv4_x and conv5_x to the previous layer.

But T=5 in pool5, with stride 2, should have the 5 connections doubly spaced from pool5 to the previous layer.

In Table 1, the receptive field of conv5_x is 17τ. I can figure this out by reading figure 3.

But, if the 5 connections from pool5 to previous layer is double spaced, the receptive field of pool5 should be 17(receptive field of conv5_x) + 4(left) + 4(right) = 25τ. Not 49τ as in Table 1.

I think I may misunderstand something. Would you pls help? Thanks.

[SeanHyenKiong]

I'm also confused by the setting of temporal CNN. The input of the temporal network is 5 temporal chunks(nFrames = 5;), how does it work? How to build the whole temporal CNN?