why split foreground and background features based importance but then combine them for next submaniford conv?

[FlyingQianMM]

Hi, these codes confuse me:

Foreground and background features are splitted firstly based the predicted importance:

https://github.com/dvlab-research/FocalsConv/blob/875e7b0c931ad7d1b2577c4c2201f228c7314a57/OpenPCDet/pcdet/models/backbones_3d/focal_sparse_conv/focal_sparse_conv.py#L212

but then combine the foreground and background features to be fed into a next submaniford conv:

https://github.com/dvlab-research/FocalsConv/blob/875e7b0c931ad7d1b2577c4c2201f228c7314a57/OpenPCDet/pcdet/models/backbones_3d/focal_sparse_conv/focal_sparse_conv.py#L216

why background features participate in next conv？

[yukang2017]

Hi, thanks for your interest in our work.

The steps are: (1) we split input features into foreground and background ones, based on the importance, (2) we dilate the foreground features. (3) we combine the dilated foreground features and the background ones.

We do not delete background features. The x_fore in the line below is the dilated foreground features.

https://github.com/dvlab-research/FocalsConv/blob/875e7b0c931ad7d1b2577c4c2201f228c7314a57/OpenPCDet/pcdet/models/backbones_3d/focal_sparse_conv/focal_sparse_conv.py#L212

[FlyingQianMM]

Thank you so much for the timely reply and detailed answer!

I get the steps are: (1) predict importance by a submaniford conv (2) filter foreground centric voxel within importance by a top-k or threshold method (3) filter important voxel in the kernel size for each foreground centric voxel by a threshold method (4) add these important voxel into foreground feature, which are your mentioned dilated foreground features

Is that right?

[yukang2017]

Yes. These are the steps.