details about point-wise refinement module about cylinder3d HOT 11 CLOSED

1: the output of up0e = self.ReconNet(up1e)

2: For different channels, conv1d or linear is used to change dimension. Addition fusion is used.

from cylinder3d.

For point-wise module, it is a symmetrical counterpart compared to MLP in cylindrical features generation (as shown in Fig.2 in original paper). Channels are also symmetrical, including 128, 64, 32.

from cylinder3d.

Thanks for your explanation, but I still have two questions:
1、From my understanding, the channels in cylindrical features generation are 64, 128, 256, so is it that in final version the channels are changed to 32, 64, 128?
2、Are the features before and after 3D convolution networks added together in point-wise refinement?

from cylinder3d.

1: It should be 256, 128, 64; not 128, 64, 32; Sorry for the mistake.

2: The features after 3D convolution is used for point-wise refinement; Note that these features should be transformed to the point-wise features first.

from cylinder3d.

I'm still a little confused about the details:

1. The features transformed to point-wise features are the output of up0e = self.ReconNet(up1e) or up0e.features = torch.cat((up0e.features, up1e.features), 1)?
2. How to fuse the features before and after 3D convolution? It seems that the channels are different(256 and 64/128), so do you use concatenation to fuse the features?

from cylinder3d.

Thanks very much!

from cylinder3d.

Hi.

It seems that the config file (config/semantickitti.yaml) is just a simplified version and not for final use.
What are the training settings details such as batch_size per GPU, num_gpus, epochs?

Thanks.

from cylinder3d.

Can you share the code of pointwise？Thank you! Email:[email protected]

from cylinder3d.

Can you share the code of pointwise？Thank you! Email:[email protected]

Did you get it or did you solve it yourself?

from cylinder3d.

1: It should be 256, 128, 64; not 128, 64, 32; Sorry for the mistake.

2: The features after 3D convolution is used for point-wise refinement; Note that these features should be transformed to the point-wise features first.

Hi, Could you share any more knowledge on how to backreference Voxel partitions into points?

from cylinder3d.

1: the output of up0e = self.ReconNet(up1e)

2: For different channels, conv1d or linear is used to change dimension. Addition fusion is used.

Hi, the dimension of up0e features in 1 is 64, so the voxel features should be transformed to the point-wise features (pointnum*64) and then use linear to change the demension to 256, then use the similar mlp to do the 256-128-64-out, is that right?

from cylinder3d.