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

Question about the loss #42

Closed Kguo-cs closed 3 years ago

Kguo-cs commented 3 years ago

In your paper, the loss is the sum of negative log-likehood of the position. But in the code, it is the sum of negative log-likehood of the relative position. I don't think they are the same because the prior position will influence the later position. Could you explain it for me? Thank you.

abduallahmohamed commented 3 years ago

Hi, Thanks for asking: The positions are in meters domain, for example pedestrian 1 is at (7.30,6.5) while pedestrian 2 is at (0.3,22.5). If you trained a deep model using these numbers you will not converge because of the magnitude of the coordinates. Thus usage of relative positions is indeed important as it transforms the coordinates into a more friendly form with small deltas all under 1. You can think of it as a normalization approach for the coordinates, eventually at the evaluation stage these relative positions is turned into absolute positions for the calculations of ADE and FDE metrics. Another way to see it that the relative positions is speed rather than positions. At the time of writing the paper we thought it as a normalization method and we formulated the problem as positions predictions as this is our final goal. Also, this approach was used by several prior and later works yet the problem still formulated as positions not relative positions in consideration that this is a normalization approach.

I hope my answer is sufficient and thanks for taking interest in my work.

On Wed, Jan 27, 2021 at 1:21 AM Kguo-cs notifications@github.com wrote:

In your paper, the loss is the sum of negative log-likehood of the position. But in the code, it is the sum of negative log-likehood of the relative position. I don't think they are the same because the prior position will influence the later position. Could you explain it for me? Thank you.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/abduallahmohamed/Social-STGCNN/issues/42, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFIVLG7QCHVLQ42ERC6MCE3S365JLANCNFSM4WUYJLGA .

--

Abduallah Mohamed abduallahmohamed.com