主要关心这个object bank是哪里来的

[algorithm3d]

1. 如果object bank是单数据集内（train或val）的，在test上提升这么明显，那这个方法确实很有价值。
2. 如果有跨数据集object或者含有test data object（正常来讲不太可能），那应该标明使用了额外的数据（add data），实际价值可能会打折，有点像跨数据集增强的思路。

[JinglinZhan]

1. 如果object bank是单数据集内（train或val）的，在test上提升这么明显，那这个方法确实很有价值。
2. 如果有跨数据集object或者含有test data object（正常来讲不太可能），那应该标明使用了额外的数据（add data），实际价值可能会打折，有点像跨数据集增强的思路。

1. For results test on nuScenes val set, the object bank is generated from nuScenes train set. For results test on nuScenes test set, the object bank is generated from nuScenes train and val set. (The same strategy that used in OpenPCDet)
2. This work does not include any cross-dataset attempts. Theoretically, exploring effective strategies for expanding the ground-truth object database, including object-level augmentation, transferring objects from external datasets, or using generative AI techniques, could further optimize the final results. However, there are still many challenges associated with these strategies, and we are in the process of addressing them.

[algorithm3d]

1. 如果object bank是单数据集内（train或val）的，在test上提升这么明显，那这个方法确实很有价值。
2. 如果有跨数据集object或者含有test data object（正常来讲不太可能），那应该标明使用了额外的数据（add data），实际价值可能会打折，有点像跨数据集增强的思路。

1. For results test on nuScenes val set, the object bank is generated from nuScenes train set. For results test on nuScenes test set, the object bank is generated from nuScenes train and val set. (The same strategy that used in OpenPCDet)
2. This work does not include any cross-dataset attempts. Theoretically, exploring effective strategies for expanding the ground-truth object database, including object-level augmentation, transferring objects from external datasets, or using generative AI techniques, could further optimize the final results. However, there are still many challenges associated with these strategies, and we are in the process of addressing them.

由于之前我有试过这种类似思路，在kitti上小目标提升还是挺明显的，其他数据集就没有试了，所以有此疑问。这样的话，确实证明了你们提出的这种数据增强方法很有效对于模型的性能，期待早日看到你们的code。