STN shouldn't be used for semantic segmentation

[sidml]

Hi.
I see that in this repo, STN is used for semantic segmentation case also.
According to the author, for semantic segmentation T-Net mayn't be able to find a canonical pose, so TNet is not recommended.
Please see this issue for more details.

[fxia22]

In the example I used it for object part segmentation so T-net is still relevant. If you use it for segmentation for point cloud in the wild feel free to adapt it and remove T-net

[sidml]

Thanks.
For object part segmentation, does T-Net seem to improve performance (compared to network without it) ?

[lllmmmwww]

Hello, can we use the train_segmentation.py to complete semantic segmentation？