About the computation of percentages of collisions

[pedro-mgb]

Hi. Thank you for making these experiments public.

I have a question regarding the computation of the percentages of collisions for social GAN, that you have presented here.

According to this, the collisions statistics are being done for each of the K samples outputted by the model, which makes the results being shown quite a bit inflated, as opposed to just using a single sample per pedestrian (e.g. the one with lowest ADE, although from what I seen it doesn't really matter).

Also, it would make sense to accompany those percentages also by the percentages of collisions occuring in ground truth data. For instance there might be cases where a 'collision' is just two people walking close to each other, which is quite likely in a densely crowded scene like Univ. Also an idea would be to only show the collisions that do not occur in ground truth data. But I digress at this point.

Lemme know if you've thought about this, and if you agree, or not. Thank you, have a good day!