Closed anthonyhu closed 3 years ago
Hi @anthonyhu thanks for the fixes, looks great! The only change to the PR I'd like to make is that the NuScenes train/val splits should be kept the same as the ones used in the paper. Unfortunately the official splits are invalid for the map prediction task since the ego-vehicle visits the same stretch of road in both train and val, allowing the network to overfit. Other than that I'm happy to accept the PR :)
Hey Tom! I see - I had changed the val split in a former commit but in this PR I'm not touching it (src/data/nuscenes/splits.py
is unchanged) :)
Ah my apologies - didn't see the commit which changed them back! Accepting the PR now :)
This PR fixes a few issue when generating the dataset labels and training the models.
In
render_polygon
, sometimes the list of polygon is empty, in that case nothing should be done right?