BDD_AD dataset and BDD_OIA dataset

[JaneFo]

Hello, thank you for such excellent work. I want to ask you a question, why not align the BDD_AD dataset with the BDD_OIA dataset, that is, jointly optimize the two sub-networks and train two tasks together?

[fyctime052]

Thank you so much for your comment and question!

In the NLE-DM network, the natural-language description is considered as the a more concise and efficient alternative of the reason to explain the driving actions. So, we do not train the network to jointly predict the reasons and descriptions. However, taking into account the correlation between the two tasks, jointly predicting them may lead to different outcomes (the prediction performance of the reason may be improved), which deserves further experimental testing.

Thank you again for your question. I hope I have clarified your doubts.

[JaneFo]

您好，我是付峥，来信我已收到，我会在查收邮件之后给您回复。