Nuscenes about oatomobile HOT 8 CLOSED

Hi @JasonMa2016,

The oatmobile.baselines should be easily adaptable to nuscenes.prediction API.

We didn't plan to release this since oatomobile is focused on CARLA (i.e., control tasks) and not that much on prediction tasks but we may take this into account! In the meanwhile, let me know if you have any trouble adapting the code and we can give you a hand. Also, take a look at PRECOG's codebase on how to preprocess NuScenes LIDAR point-clouds.

from oatomobile.

Thanks for the pointer! I wonder if OATMOBILE's (unreleased) code base for nuscenes utliizes PRECOG's codebase for preprocessing NuScenes LIDAR point-clouds?

from oatomobile.

Yes! But only the LIDAR preprocessing methods, we didn't use any of the multi-agent modelling PRECOG does!

from oatomobile.

I see, thank you! I wonder if the architecture detail and training hyperparameters for the Nuscenes version of your methods can be shared?

from oatomobile.

It is the same one used in this repo, by adapting the output shape and input modalities accordingly.

from oatomobile.

Thanks for the response! Sorry to bother again, but I noticed the OATOMOBILE utilizes features such as "is_at_traffic_light" and "traffic_light_state" in computing the visual features for the models. I wonder if those information are used in the NuScenes versions of the models, and if so, how are they extracted? I am having trouble finding relevant code in PRECOG's codebase. Thank you!

from oatomobile.

No, we just use the LIDAR observation for the NuScenes experiments.

from oatomobile.

Thank you! These are all the questions I had.

from oatomobile.