Questions about Uni3D.

[kellenf]

1. Use feature copy is not elegant in deployment. If I use 5 dataset, need copy 5times which cost much memory and increase latency.
2. The Semantic-level Feature Coupling-and-Recoupling Module is too fancy, and it improve little performance compared with domain-attention module.I means this module only useful in paper not in industry?
3. This new BN layer can reach same latency performance as the original BN layer?

[BOBrown]

Thanks for you questions.

A1: the feature copy method is used after obtaining the 3D and 2D features, rather than the original points. Actually, we did not consider the situation that there are a lot of datasets (e.g., 5 datasets) to be merged to perform the joint training, which is a reseach direction we are conducting. A2: The domain-attention module designed in the 3DTrans repo is an alternative option for the Semantic-level Feature Coupling-and-Recoupling Module, if you want to perform the multi-dataset joint training (perhaps it will have lower performance than C.R. module). But, the domain-attention module is indeed an effective method for converting the dataset-shared BEV_features into the dataset-specific BEV_features, and we conducted the corresponding experiments and found its effectiveness (especially for the train-and-test consistency) for the multi-dataset joint training, and thus, the main code of implementing the domain attention is released in 3DTrans-v0.1.0. A3: The proposed BN layer has the same latency performance as the original BN layer during model inference.