search

abduallahmohamed / Social-STGCNN

Code for "Social-STGCNN: A Social Spatio-Temporal Graph Convolutional Neural Network for Human Trajectory Prediction" CVPR 2020
MIT License
483 stars 141 forks source link

anorm problem about step_rel and step_ #57

Closed Ian-Shih closed 2 years ago

Ian-Shih commented 2 years ago

I have seen #15, but I still can't understand it. If people A and B don't move, both their step_rel=(0,0), use step_rel to calculate L2 distance means that A and B are closest? Because (0-0)^2+(0-0)^2=0 . I don't think the distance about velocity has a physical meaning .

My English is bad if you don't know what I ask, please let me know, thanks.

abduallahmohamed commented 2 years ago

Thanks for your question! We used 1/L_2 of difference in distance to represent on the relationship between the agents. If agents are far away it will have a vey small weight and the closer the bigger the value of the relationship becomes. The special case when the L_2 is zero, so 1/0 = inf = the strongest relationship. We set this case to be 0, the reason behind this selection that setting any value beside 0 for this situation as I recall lead to a degraded performance. Reflecting on this, it’s like two persons walking in parallel next to each other which can be seen globally as a one entity. So, setting to 0 allowed us to disentangle this special case and treating it as a single entity in concept even though they are several entities.

I hope this helps. Feel free to contact me if you have any questions.

Best, Abduallah

On Thu, Mar 10, 2022 at 12:05 AM Ian-Shih @.***> wrote:

I have seen #15 https://github.com/abduallahmohamed/Social-STGCNN/issues/15, but I still can't understand it. If people A and B don't move, both their step_rel=(0,0), use step_rel to calculate L2 distance means that A and B are closest? Because (0-0)^2+(0-0)^2=0 .

My English is bad if you don't know what I ask, please let me know, thanks.

— Reply to this email directly, view it on GitHub https://github.com/abduallahmohamed/Social-STGCNN/issues/57, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFIVLGYXFJ3CQSCQATBU4LTU7GULZANCNFSM5QL2ZXPA . You are receiving this because you are subscribed to this thread.Message ID: @.***>

--

Abduallah Mohamed abduallahmohamed.com

Ian-Shih commented 2 years ago

Thanks for your reply, but I don't focus on being 0 or not being 0. Maybe my bad example made you confused. Let me show another example: image In my opinion, replacing Steprel with Step is better, but the result is worse, just like #15.

abduallahmohamed commented 2 years ago

I agree with you that step_ might be better than steprel but the results says otherwise. It could be the importance of the deltas vs the importance of displacement, and maybe the acceleration will work better? or even having edges with step, step_rel & step_acc ? Overall, it's an assumption from our point of view and the model might says something else driven by the data and the nature of data representation.

Ian-Shih commented 2 years ago

OK, I'll try more ideas about this issue. Thank you.