Map based LSTM

[Cram3r95]

Hi guys!

These baselines are really wonderful to start with Motion Forecasting. In my particular case, I am trying to include map information to the LSTM in order to compute multimodal predictions that are directed towards plausible lanes (computed using map_utils.py). However, I have not seen in the train function that the input to the encoder is influenced by the lanes:

But, on the other hand, in the infer_map function, you specify some CANDIDATE_CENTERLINES and modify the input of the encoder:

Do you use these map lanes only during inference or also while training?

[jagjeet-singh]

Hi! The map is used during training as well. Instead of using all candidate centerlines, we use just the oracle centerline. The input features in the above "train" code are for a pre-selected oracle centerline. Refer the feature computation here for more details. Hope this makes sense.

[Cram3r95]

then you train with just the oracle (the groundtruth centerline) and in the inference you do the same for each candidate, so all map information is not included for each iteration of the for loop, but only each candidate, right? And what happens if I want to predict k=6 trajectories (required for the leaderboard) but I have less than 6 candidates in front of me?

By the way, do you encode both the past and future information of the centerline or just the future?