search

autonomousvision / transfuser

[PAMI'23] TransFuser: Imitation with Transformer-Based Sensor Fusion for Autonomous Driving; [CVPR'21] Multi-Modal Fusion Transformer for End-to-End Autonomous Driving
MIT License
1.04k stars 175 forks source link

About the expert in the result #51

Closed Watson52 closed 2 years ago

Watson52 commented 2 years ago

Hi, I want to know that the expert with so high score in the result is because of its privileged information? I guess the expert is the agent from the first stage of the LBC.

ap229997 commented 2 years ago

Yes, expert uses privileged information but it is different than the first stage (teacher model) of LBC. Expert uses A-star planner and collision avoidance heuristics to drive.

Watson52 commented 2 years ago

I see. Well, what if the expert did not have privileged information? Or what do you think of the result with the A-star planner and collision avoidance heuristics?

ap229997 commented 2 years ago

The current expert is still not perfect and suffers some infractions since there are certain loopholes in the heuristics.

Watson52 commented 2 years ago

Thanks! I see.

Watson52 commented 2 years ago

Hi again. I try to realize the LBC but I didn't quite get it especially the 'topdown'. As you say that the dataset you provided could not train the LBC because of topdown. Could you please tell me what does it mean? Your repply with be appreciated.

ap229997 commented 2 years ago

We have also provided a large scale dataset which contains topdown and you can use it as input to train the teacher model of LBC. Here, topdown means topdown semantic segmentation of the scene. You can also look at LBC repo for more details on how to train the model.

Watson52 commented 2 years ago

We have also provided a large scale dataset which contains topdown and you can use it as input to train the teacher model of LBC. Here, topdown means topdown semantic segmentation of the scene. You can also look at LBC repo for more details on how to train the model.

I see. Thanks a lot.