search

mmahdavian / STPOTR

Human Pose and Hip Trajectory Prediction Using Transformers
GNU General Public License v3.0
11 stars 2 forks source link

Hello, is there another way to download, I can't get on Google, thank you very much! #2

Closed kstudy123 closed 1 year ago

mmahdavian commented 1 year ago

Hi. Which file can't you download?

kstudy123 commented 1 year ago

Hello, sorry I just saw your reply, thank you very much, the problem has been solved. But I have one more question: about your internal non-autoregressive model, I checked the potr but found that it is not explained in detail, may I ask if the input qkv of the internal decoder of the model is only summed with the position code, and how it achieves non-autoregressive? Thank you very much for your patient answer!

mmahdavian commented 1 year ago

Hello again. Sorry for delay in my reply. The model adds the encoder embeddings for trajectory with the output of multi-head attention module which is applied to encoder embeddings of pose and trajectory sections. Then it is used in another Multi-head attention module as q and k. The v would be the last hip position coppied m=20 times. The decoder actually tries to find the difference between that point and the next values in each frame. The model becomes non-autoregressive for that reason. We don't need to calculate each value one by one. The model tries to predict them all together at the same time.