search

umautobots / bidirection-trajectory-predicter

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

Question - Dataset handling #3

Closed ksachdeva closed 3 years ago

ksachdeva commented 3 years ago

Hi @MoonBlvd

I find myself a bit confused with the dataset input pipeline in your experiment. I have ideas about what is happening but would appreciate it if you could verify or correct my understanding.

Typically in my own experiment, I prepare the trajectory dataset (for e.g. ETH/UCY) by slicing the trajectories of each pedestrian into 20 points (x,y) [first 8 become obs and rest 12 are to be predicted). The slicing is done in a rolled fashion with a window of 1. I then store these slices in npz or pickle format.

In your source code, I see the notion of "curr_pos" and "first_history_indices". The most I can guess here that what I do and store in files you are doing it dynamically at each batch during training. The naming "first_history_indices" is not necessarily intuitive so I am not sure.

Would appreciate it if you could throw some light on it.

IMHO, preparing the 20 point trajectories will make the experiment faster and the code will be simpler but I am assuming that you have inherited this approach from 'trajectron-plus-plus' where they take social interactions into account.

Regards & thanks Kapil

MoonBlvd commented 3 years ago

Hi @ksachdeva , you understanding is basically correct. We actually used trajectorn++ data preparation and only added/removed outputs of some of their functions. We trained and tested all models with 8 history frames and 12 future frames, same to Trajectron++.