Paper understanding

[saharshleo]

Hello,
What is the exact dimensions of input to deformable transformer encoder?
From what I understood:

• Input video is sampled for T frames
• Extract frame features using TSN/C3D {x^t}_t=1^T
• Then L convolutional layers for each frame feature giving {f^l}_l=1^L

So the input is of TxL temporal dimension right?

[ttengwang]

Hi, the $L$ convolutional layers all have a stride K=2, so resultant temporal dimension is $T + T/K + ... + T/K^{L}$