search

cunjunyu / STAR

[ECCV 2020] Code for "Spatio-Temporal Graph Transformer Networks for Pedestrian Trajectory Prediction"
MIT License
338 stars 78 forks source link

External Graph Memory #10

Closed MazenHamdi closed 3 years ago

MazenHamdi commented 3 years ago

Hey, great work. I have a couple of questions, currently I am trying to compare different architectures by changing the original STAR architecture (for example deleting the 2nd encoder, changing the number of heads/layers..) and now I want to get rid of the graph memory and see how it works, I want to know if GM is essential for the temporal aspect of the transformer? and is it even possible to talk about getting rid of the GM. Thanks in advance

cunjunyu commented 3 years ago

Hi, thanks for your interest in our work.

GM serves to retain the effect imposed by those pedestrians who have left the scene but used to be present. Some workarounds may do the same thing. For example, regarding the pedestrian as a static point the moment when they disappear from the scene.

Thank you.

MazenHamdi commented 3 years ago

Oh I see, thanks a lot, so I tried to compare the model with and without a GM, here are the results: with GM: ADE/FDE: 0.160/0.312 without GM: ADE/FDE: 0.129/0.231

does it make sense that the results without the GM are better ? or maybe because I didn't run enough training samples (because the models are stochastic)

cunjunyu commented 3 years ago

does it make sense that the results without the GM are better? or maybe because I didn't run enough training samples (because the models are stochastic)

GM does not provide a very significant improvement. Because the model is stochastic, you may get a different result in each run. The reason why we use GM is that GM brings a slight improvement in the deterministic version.

Thanks.