Temporal Graph Neural Networks have been used for motion prediction (or traffic forecasting) in the following recent papers:

Dynamic Multiscale Graph Neural Networks for 3D Skeleton-Based Human Motion Prediction Graph Neural Network for Traffic Forecasting: A Survey

I was wondering, how to use them or any of the available temporal graph neural networks for motion prediction when the graphs don't have the same number of nodes / edges?

Let's say I have recorded a motion of a skeletons that have 5 joints and 10 joints. Then, after the model is trained, I want to use that model for a motion prediction of skeleton with 7 joints. Each node (joint) has 3 attributes, coordinate x, y and z.

Is that even possible, or is there trivial way how to do that? Could something like graph padding (creating nodes without any edges) work in that case?

For instance, in the first paper when they initialize temporal attention or spatial attention layer, it depends on the number of nodes. It creates a learnable weight matrix with dimension N, therefore if we change the input function (i.e. inference) it fails due to impossible matrix multiplication.

  • $\begingroup$ Can you please put your specific question in the title? "Temporal Graph Neural Network for motion prediction" is not a question. $\endgroup$
    – nbro
    Feb 10, 2022 at 20:52


You must log in to answer this question.

Browse other questions tagged .