search

HaochengWan / PVT

PVT: Point-Voxel Transformer for 3D Deep Learning
https://arxiv.org/abs/2108.06076
MIT License
87 stars 8 forks source link

There are some bugs #5

Open Starak-x opened 2 years ago

Starak-x commented 2 years ago

Hi, it's a great work. But I find some bugs in it. rel_pos should be fed to point_features in PartPVTConv and SegPVTConv, in pvtconv.py

2022-04-28 14-50-45 的屏幕截图 2022-04-28 14-46-50 的屏幕截图 there are also bugs in shared_transformer.py SA_Layer use rel_pos,but Shared_Transformer don't receive it and don't feed it to SA_Layer 2022-04-28 14-57-43 的屏幕截图 2022-04-28 14-47-32 的屏幕截图

the code above are screenshot in github code, as follow: 2022-04-28 15-00-08屏幕截图

Starak-x commented 2 years ago

Maybe they can be modified as follow: In PVTConv.py: 2022-04-28 15-01-54 的屏幕截图 In Shared_Transformer.py: 2022-04-28 15-02-56 的屏幕截图 Maybe the code you send is not the final version, the motivation is wonderful.

zouwenqin commented 2 years ago

I also find these bugs, do you test the final result? I modified as you and used it for other task, but I can't got a good result.

tomanick commented 2 years ago

@zouwenqin I also found these bugs. Afterwards, I tried to fix these bugs by myself and I could run PVT for training, but I could not reproduce the results in the paper. When I wanted to train on my own dataset, I found that the result was worse than PVCNN's result.