point cloud head problem

[gaojh0333]

Thank you for your excellent work.

I noticed that enable_point=False is used in the released model, and the point head weights do not seem to be included in the open-source checkpoint. Instead, the reconstruction appears to rely on the predicted camera and depth outputs.

Could you please explain the reason behind this design choice? Was the point head disabled because it did not improve performance or generalization?

Also, would it be possible to share the pretrained weights for the point head as well? I would be very interested in experimenting with that branch.

Thank you!

[LinZhuoChen]

Hi~ Thank you for your attention.
We only train point head in the first stage, and then we remove point head in the second stage due to high memory cost and no significant improvement.

[gaojh0333]

Hi~ Thank you for your attention. We only train point head in the first stage, and then we remove point head in the second stage due to high memory cost and no significant improvement.

Could you please explain the processing of ground truth point clouds during point cloud monitoring? It's unlikely that the data is directly reconstructed from the ground truth values ​​of camera intrinsics and extrinsic parameters and depth. Does the depth and camera endothelial/extrinsic parameter ground truth values ​​undergo any processing?