A question about the result of 3DRegNet.

[QWTforGithub]

Hi, thank you very much for your well-organized pointDSC code. I noticed the experimental result of 3DRegNet on 3DMatch in your paper. Would you like to provide the code of 3DRegNet on 3DMatch? Thank you very much!

[XuyangBai]

Hi, thanks for your interest. As shown in our paper (Supplementary 7.2) 3DRegNet does not work well on 3DMatch with their registration block so I only use the classification block to produce the inlier probability. Then 3DRegNet becomes a variant of PointCN (Learning to find good correspondence paper) which is the same as the following module (but you should uncomment the InstanceNorm first). I believe you can quickly derive a 3DRegNet model and use our implementation to train it by yourself.

PointDSC/models/PointDSC.py

Lines 55 to 62 in b009d53

	for i in range(num_layers):
	layer = nn.Sequential(
	nn.Conv1d(num_channels, num_channels, kernel_size=1, bias=True),
	# nn.InstanceNorm1d(num_channels),
	nn.BatchNorm1d(num_channels),
	nn.ReLU(inplace=True)
	)
	self.blocks[f'PointCN_layer_{i}'] = layer

[QWTforGithub]

Thank you very much for your reply. I'd like to confirm again:

1. Uncomment # nn.InstanceNorm1d(num_channels)
2. comment Self.blocks [f'NonLocal_layer_{I}'] = NonLocalBlock(num_channels)
   Should I do this?

[XuyangBai]

Yes, that setting is what I used during my experiment.

[QWTforGithub]

Hi, thanks for your interest. As shown in our paper (Supplementary 7.2) 3DRegNet does not work well on 3DMatch with their registration block so I only use the classification block to produce the inlier probability. Then 3DRegNet becomes a variant of PointCN (Learning to find good correspondence paper) which is the same as the following module (but you should uncomment the InstanceNorm first). I believe you can quickly derive a 3DRegNet model and use our implementation to train it by yourself.

PointDSC/models/PointDSC.py

Lines 55 to 62 in b009d53

	for i in range(num_layers):
	layer = nn.Sequential(
	nn.Conv1d(num_channels, num_channels, kernel_size=1, bias=True),
	# nn.InstanceNorm1d(num_channels),
	nn.BatchNorm1d(num_channels),
	nn.ReLU(inplace=True)
	)
	self.blocks[f'PointCN_layer_{i}'] = layer

And, I only use classified loss to train the network, i.e. BCE loss in pointDSC?

Yes, that setting is what I used during my experiment.