search

Haiyang-W / DSVT

[CVPR2023] Official Implementation of "DSVT: Dynamic Sparse Voxel Transformer with Rotated Sets"
https://arxiv.org/abs/2301.06051
Apache License 2.0
353 stars 28 forks source link

Errors included in the process of assigning the Position embedding vector for multi-head self attention #50

Closed junhyung-SPALab closed 10 months ago

junhyung-SPALab commented 11 months ago

Thanks for your work. In the process of studying your great open source, I leave a question.

There seems to be a bug in the process of using the position embedding vector in the operation of the DSVT Block.

It seems that the position embedding vector required for Attention is assigned incorrectly.

The questions have been summarized in the image below. ////////////////////////////////////////////////////////////////// This is an example of a case where DSVT is composed of one stage and each stage is designed with two blocks.

bug_0 bug_1

//////////////////////////////////////////////////////////////////

Thanks.

chenshi3 commented 11 months ago

Thank you for raising this issue! We have also identified this bug and documented it in OpenPCDet. The reported results were obtained from this version, and we have chosen not to rectify the bug in the code. Furthermore, I attempted to address this bug but noticed negligible performance enhancements. I will make a note of this bug in the Readme for clarity.

junhyung-SPALab commented 10 months ago

Thanks for confirming my issue. I think your README update will be helpful to other researchers in the future.