Multiple view overlaying problem.

[BHC1205]

Thank you for your great work.

1. I was wondering if the overlap area here is not specially treated, but simply covered with the information of the latter view. Is it because sum or avg doesn't work well, or simply because it wasn't implemented?
   https://github.com/fudan-zvg/DeepInteraction/blob/main/projects/mmdet3d_plugin/models/utils/decoder_utils.py#L758

2.Is it because of its superior performance that DynamicConv is used to replace the cross-Attention module in the decoder?

[Alexander0Yang]

1. I was wondering if the overlap area here is not specially treated, but simply covered with the information of the latter view. Is it because sum or avg doesn't work well, or simply because it wasn't implemented?
   https://github.com/fudan-zvg/DeepInteraction/blob/main/projects/mmdet3d_plugin/models/utils/decoder_utils.py#L758

Yes, queries in overlap area only keep the information from the latter view in the current implementation, because it is simple and the number of queries in overlap area is negligible. Using information from multiple view will be helpful by intuition, but we haven't tried.

2. Is it because of its superior performance that DynamicConv is used to replace the cross-Attention module in the decoder?

Yes, replacing vanilla transformer decoder layer with our predictive interaction layer for both modalities can bring better performance as shown in Table 3(a) of our paper.

[BHC1205]

Thank you for your detailed response.

[BHC1205]

I'm sorry to reopen this issue, but there are still some doubts: I tried using DynamicConv compared to the vanilla transformer, and it is indeed more effective, but I still don't understand why it is effective. Is it because the task being processed is RoI Feature, and I guess vanilla transformer may have difficulty describing its position information? If it is convenient, can you explain the reason why DynamicConv can bring better performance? Thank you.