amazon-science / explainable-trajectory-prediction

Official code repository for the ICLR 2022 paper "You Mostly Walk Alone: Analyzing Feature Attribution in Trajectory Prediction".
https://openreview.net/forum?id=POxF-LEqnF
Apache License 2.0
13 stars 4 forks source link

Neighbors' Radius vs Attention Radius #11

Closed m43 closed 1 year ago

m43 commented 1 year ago

Hi Osama,

Thanks for the nice work!

I was wondering why you had separated the neighbors' radius from the attention radius. The two are, for example, defined here. For what I could dig out, the neighbors' radius is used to compute the scene graph here, and the attention radius is used to overwrite the position standardization parameters here and again here when recomputing the scene graph.

Should the two parameters be separated? Would you remember why you had separated them? Would you know how to configure the radii when adapting the model for a new dataset?

Best, Frano

os1a commented 1 year ago

Hi Frano,

Thanks for reaching out regarding our paper.

As for the two parameters, the neighbors radius control how many neighbors are needed to compute the scene graph (as you found out). In the Sport-VU dataset, this corresponds to all players. However, in the original Trajectron++, they used smaller values (to compute the grapy only for a sub-set of neighbors for crowdy environments).

The other parameter controls the distance to the neighbors to be considered when computing the scene graph. So this can be seen as a way of controlling the number of neighbors (the closer to be used). These two parameters come from the original Trajectron++ and we used them as such. More information can be found here: https://github.com/StanfordASL/Trajectron-plus-plus/blob/ef0165a93ee5ba8cdc14f9b999b3e00070cd8588/trajectron/environment/scene_graph.py#L166

Best, Osama

os1a commented 1 year ago

I am closing this, please feel free to reopen it if you have more questions.