Code for other baselines

[HankKung]

Hi wondering when will the code for baselines be released:

• traditional stagewise method
• PnPNet-Vision

Since the results that compare PnPNet-Vision and ViP3d provided by ViP3D and tab. 5 in UniAD are contradictory. PnPNet-Vision outperforms ViP3D by a large margin in their paper. I believe it is quite urgent that the authors should release the full code for the implementation.

Also, I notice that the authors do not plan to release the goal-based and heatmap-based ViP3D because they are not in good shape.
However, it will be great and crucial to reproduce the experiments in the paper exactly.

[HankKung]

Hi @HankKung

Does your problem solved ?

Nope, have not received any updates.

[GentleSmile]

It is unfair to compare different prediction methods using different detection backbone. PnPNet-Vision in UniAD is reimplemented with BEVFormer which has strong detection performance. The performance of end-to-end prediction heavily depend on detection performance because of error accumulation.

[GentleSmile]

Sorry for my delayed response (one year later) to your question for performance comparison between PnPNet-Vision and ViP3D, because I only see the first version of this issue and did not received a notification after you edited the issue.