Explain about paper

[LeVan146]

Hi authors, I'm reading your paper and I see the result very good. But I still not clear about some information in paper:

• In the above image, Can you tell me what axis are, and based on my knowledge, deep features is 1D vector, so what formula you use to convert 1D vector to a number in [-0.15;0.1].

• Why content motion is shape of signal and style in scale and bias of signal?

• Can you explain more about why he temporal 1D convolution return the unit quaternions tend to be smooth? Based on my knowledge, I just know temporal 1D convolutonal network is famous for time-series, but I can't find the reason why it return continuites number? It is fine for me if you just put a reference to explain this ^^

I would be grateful if you can explain my concerns to make me more understand about your great work.

[HalfSummer11]

Hi, thanks for your interest in our work! Regarding your questions:

• As the caption suggests, we are visualizing a specific channel of the 1D vector deep feature, so it's a number. Here the y axis is the temporal axis. (Basically, we have a feature matrix C T, where C is the number of channels and T is the temporal length. We visualize one row 1 T.)
• This is an intuitive explanation of what the network does, just as image style transfer works claim/assume the Gram matrix/the second-order statistics correspond to the image style. For input/output signals, each channel corresponds to the x/y/z coordinate of a specific joint. The shape of the signal tells us how this joint moves in space (e.g. the wrist joint may move back and forth during walking), and the bias/scale of the signal gives the range of the joint motion, which is more related to style, e.g. a big motion range suggests an energetic style. (see our video for a visual explanation). Furthermore, in our framework, we explicitly inject style information by manipulating the scale/bias of the signal, while the shape of the signal comes from the content code. This is the way we model the problem.
• We think this more of our empirical experience with 1D convolutions on motion data -- a "smoothing kernel" is often learned, leading to continuous outputs, e.g. footskate artifacts.

[LeVan146]

Thank a lot for your explanation ^^