search

umautobots / bidirection-trajectory-predicter

The code for Bi-directional Trajectory Prediction (BiTraP).
Other
78 stars 23 forks source link

Question about data pre-processing #11

Closed d-zh closed 2 years ago

d-zh commented 2 years ago

Hi, Your work is impressive! I have some questions about JAAD/PIE dataset pre-processing. Obviously, JAAD/PIE use the bounding box (x1,y1,x2,y2) as input. I noticed that your codes use {'normalize_bbox': True} (see datasets/JAAD.py line 34), which means the input bounding boxes are the displacement of two adjacent frames. So, I wonder the loss function is "y_pred - y_gt" or "delta_y_pred - delta_y_gt" ? The first one is the original trajectory and the last one is the displacement of adjacent frames. And which is your evaluation way? Thank you very much!

MoonBlvd commented 2 years ago

Hi @d-zh , the normalization is to normalized the box to zero-one or plus-minus-one. I believe you fill find how the loss function is computed in the loss code here and here. if you trace back, you will find what is the y_gt and y_pred.

d-zh commented 2 years ago

Thanks a lot, I got related codes.