Question about the 3D PE in Qformer-3D

[moonsbird29]

Thanks for your great work!
I am not an expert in autonomous driving and mainly work on other research topics, but I am still very interested in your work, especially the implementation details of Qformer-3D and the 3D positional encoding. However, I could not find the corresponding implementation in the codebase.
I would appreciate it if you could point me to the exact Python file where it is implemented. Thank you!

[Whale-ice]

same question ， can‘t find q-fomer codes

[moonsbird29]

same question ， can‘t find q-fomer codes

Hi, the implementation of 3D PE in PETR/PETRv2 are more clear to understand. (in PETR/projects/mmdet3d_plugin/models/dense_heads/petr_head.py). I hope this can helps you!

I also curious about the impact of Qformer-3D initialization, what will happend if not init by pretrained 2D Qformer. The Qformer-3D in ORION (https://arxiv.org/abs/2503.19755) is trained from scratch.