Which paper is torchvision.ops.deform_conv2d from?

[songyuc]

📚 Documentation

I want to know which paper torchvision.ops.deform_conv2d is from, is it DCNv1 or DCNv2?

[datumbox]

@songyuc On latest master both! Check out this one:

vision/torchvision/ops/deform_conv.py

Lines 23 to 28 in 3d60f49

	Performs Deformable Convolution v2, described in
	`Deformable ConvNets v2: More Deformable, Better Results
	<https://arxiv.org/abs/1811.11168>`__ if :attr:`mask` is not ``None`` and
	Performs Deformable Convolution, described in
	`Deformable Convolutional Networks
	<https://arxiv.org/abs/1703.06211>`__ if :attr:`mask` is ``None``.

The v2 version was added recently.

[songyuc]

@datumbox , is there an implementation of DCNv2 using this op?

[datumbox]

Unfortunately we don't have an implementation of the DCNv2 network at TorchVision at the moment.

BTW feel free to reopen the ticket if you feel I closed it prematurely. :)

[songyuc]

@datumbox , do you have a plan to add it in torchvision?

[datumbox]

We don't have immediate plans but it's worth adding it to this ticket #2707 where we keep track requests. Or you can raise a separate issue to propose its addition. :)