Is it possible to support 3D convolution? about deformable-convolution-pytorch

[1zb]

The forward of ConvOffset2D contains two part: deformable_im2col and gemm. To generalize it to 3D version, the function deformable_im2col needs a rewrite (deformable_im3d2col).

Similarly, in backward propagation, following functions should be implemented:

1. deformable_col2im3d_coord
2. deformable_col2im3d

In the near future, I may take a try, but don't expect too much.

[bkvie]

What would 3d conv change, isn't the offset applied all layers anyways?

[jmarsil]

Has anyone given this a shot yet?

[KrakenLeaf]

Hi, you can take a look at: https://github.com/KrakenLeaf/deform_conv_pytorch
Let me know if you find any bugs :)

[jmarsil]

@KrakenLeaf awesome, awesome, awesome, I'll integrate this into one of my auto segmentation pipelines and let you know!!

[salmank255]

Hi @KrakenLeaf,
The 3d deformable convolution is amazing. What about 3d deformable RoI pooling? Do you have the 3d version of pooling layer? if yes please share it.

Thanks

[KrakenLeaf]

Hi @salmank255 ,
Thanks, but unfortunately I did not implement 3D deformable ROI pooling yet.
But this is definitely something to look at.

[salmank255]

Yeah @KrakenLeaf, I am working on it, but didn't complete it yet. If you convert it, please let me know.
Thanks