abduallahmohamed / social-implicit Goto Github PK

Can you please clarify the implementation of geometric angle losss at

In the figure provided at Social-Loss.png, it is shown angle b/w two edges, however from the code, it looks like the code computes the angle at the origin between any predicted points t and j.

Can you please clarify further on G-angle loss about its inputs and what angle are we computing? and can you also please tell if V_pred and V_target at [line 88] (

We found that the effect you achieved on the SDD data set is significantly better than other methods, but why are the results of the same SDD data set so different?

Can the results of the two be compared uniformly? Hope you can clear up my current confusion.

I wish you have a nice day！

In model.py line 149
As the v.shape is [batch,2,obs_len,num_ped], then (v.permute(0, 3, 1, 2)[0, :, :, 0]).shape is [num_ped,2], which means the location of num_ped pedestrians at the first time step. The location of every pedestrian at the first timestep is (0,0). I am confused that how the line 149 in model.py can obtain the max speed change for every sequence.

Can you explain how to obtain group’s maximum change of speed and cluster the motion of pedestrians?

Thanks for providing the implementations of AMD and AMV.

I am confused by the code about AMV when taking the average. In the function calc_amd_amv of the file amd_amv_kde_metrics.py, the estimated covariance matrix is averaged over all time steps and all agents. Then, the maximum eigenvalue is computed on the averaged matrix. The calculation is inconsistent with Eq. (4) in the paper. Eq. (4) says the covariance matrix is estimated for each agent at each time step, from which a maximum eigenvalue is computed. The final AMV is obtained by averaging these eigen values. I belive the two quanties, i.e., maximum eigen value of an average matrix and the average of maximum eigen values, are generally not equal.

Could you explain the difference? It will be valuable for me when applying the code to new predictions on other datasets.

Great work! Thank you for contributing the code. I am now trying to make a visualization following the provided code. But when I run the Align.ipynb, the return value in sections 4 to 7 are (181, 181, 364, 181) instead of (2253, 2253, 2356, 2253). But section 3 (Trajectorn++) returns the same value as 364. Is there any problem with my package version or the downloaded data?

Thank you very much for your great contribution. I am very interested in the results of model visualization in your paper. I want to visualize my model. What should I do?

Great work! Thank you for contributing the code. When I use python 3.8 to install it， I find that scikit_learn==1.0.1 is not match with scipy=0.19.1. I guess that your requirements.txt is wrong.

In Table 1, the AMD/AMV are given for several models. Can you share the code about how to obtain AMD/AMV for these models, like Trajectron++?

I want to generate outputs for a different datapoint/sequence and visualize it. How can I generate the predictions from the model in the format that the align.ipynb notebook takes? I am following the test scripts but cant seem to understand it.