About seed pose in systhesize

[Symbolzzz]

Very good framework.

I noticed the description of seed pose in the article:

the use of seed poses helps to make transitions between consecutive syntheses smooth. Seed poses di=1, ...,4, the last four frames of the previous synthesis, are concatenated with the feature vector for the early four frames of the next synthesis as (fi , di ), and an additional bit is used to indicate the presence of a seed pose.

But I found in the code that it seems that only the first four frames of the current GT are used as seed poses.

I want to know how to splice the predicted long sequence into a complete long sequence after slicing them into short sequences of 34 frames for a long input sequence. I am currently referring to your code framework in my work, but I encountered the problem of incoherent generated gestures when splicing the predicted short sequence of 34 frames. Is this related to the seed pose or the splicing method?

I would be grateful if you could answer my question.

[youngwoo-yoon]

Hello,

Please refer to here. https://github.com/ai4r/Gesture-Generation-from-Trimodal-Context/blob/master/scripts/synthesize.py#L121 It demonstrates motion synthesis for a long speech.

[Symbolzzz]

Thank you for your reply. It helps me a lot.🙏

[Symbolzzz]

Another question, I referred to the smooth transition method in the code, and I found that the synthesized short sequence fragments still appear incoherent. Is there a good solution for this?

[youngwoo-yoon]

It was a faily simple transtion smoothing, and there is room for improvement. I don't have a good and easy solution, but you can try adjusting the hyperparameters of generation length or number of seed poses for better transition.

[Symbolzzz]

Thanks again for your patient reply!