In the train,true sequece as input(20),why?
junjun520416 opened this issue · comments
in the article,from time tobs+1 to tpreds,the author use the predicted position from the previous social_lstm cell in place of the true coordinates, but in your train, i can't find it, you just use 20 true coordinates, and in the your test, same problem
I have tried both. For the prediction part, I have used predictions made in observation time but I have been finding out that instead of prediction, I have got the better results using true coordinates for prediction task. You can switch easily by changing the type of list.
thanks for your great answer.
another question,in the calculation grid, if the other people don't exist in the current frame,how calculate. In your code,you set the the other people to zero, but if zero exactly is in grid range, how should we do.(my english is pool, Hope you can understand.)
I could not understand what is the question exactly however, we are setting all peds that not in current windows to zero in the grid, therefore, their state will not contribute to other neighbor peds.
I mean that you set the coordinates of the person who did not appear in the current time to 0, if this coordinates is in the range of neighborhood_size, it will set ped that not in current windows to 1 in the grid
The grid has nothing with real positional coordinates. It is a mask. So basically, we are accumulating the neighbor peds for a ped by setting the grid value 1. If a ped is not in a grid of another ped, relevant cell set to 0.
@junjun520416 hi, I finished the train days ago , and could we have a discuss ??
Look forward to your reply !
@junjun520416 hi, I finished the train days ago , and could we have a discuss ??
Look forward to your reply !
my wechat: junjun520416
I have tried both. For the prediction part, I have used predictions made in observation time but I have been finding out that instead of prediction, I have got the better results using true coordinates for prediction task. You can switch easily by changing the type of list.
maybe, it's because you predict the second point by the first input(0, 0),but it's wrong because (0, 0) can predict nothing, and the second point is deadly to last point's prediction...so, i think the loss should be caculated beyond the second point.(my english is poor, sorry, i hope that you can understand me seriously)
I have tried both. For the prediction part, I have used predictions made in observation time but I have been finding out that instead of prediction, I have got the better results using true coordinates for prediction task. You can switch easily by changing the type of list.
maybe, it's because you predict the second point by the first input(0, 0),but it's wrong because (0, 0) can predict nothing, and the second point is deadly to last point's prediction...so, i think the loss should be caculated beyond the second point.(my english is poor, sorry, i hope that you can understand me seriously)
maybe should beyond the third point because two points don't have direction information.