question about the reconstruction loss during training

[Frog-2915]

Thank you very much for your excellent work on P2C ！
When training under unpaired setting or single partial setting,
how is the reconstruction loss (rebuilding the complete shape) specifically calculated?

In the unpaired setting, since incomplete and complete shapes are not matched,
is the reconstruction loss computed between the network output and the input partial shape itself?

In the single partial setting, since there is no corresponding complete ground truth,
is the reconstruction loss still computed based on the input partial shape itself？

I would greatly appreciate if you could clarify how the reconstruction loss is handled in these two cases.
Thanks again for your time and your contribution to the community!

[CuiRuikai]

Thank you so much for your interest!
I believe the reason why you have such a confusion about the reconstruction loss computation is the misunderstanding about how we formulate the unpaired completion setting.
No matter in the unpaired or single partial setting, we just regard the input as a shape with unknown incompleteness (when there is no incompleteness it means a complete shape).
So, the two setting are unified, and the loss are computed in the same manner as we specified in the paper.
That is, we divide the input shape as three-fold, one fold for encoding, one fold for completion loss, and one fold for latent reconstruction.
To turn the single partial setting as unpaired setting, what we did is as simple as copy the complete shapes into the partial shape folder, and no extra modification in the code is needed.
Let me know if you have further concerns.

[Frog-2915]

Just to confirm my understanding:

Do you mean that the reconstruction loss is computed by using the predicted point cloud to fit a certain local region of the input point cloud (rather than fitting the entire complete shape)?

When you mentioned "one fold for encoding, one fold for completion loss, and one fold for latent reconstruction",
does it mean that:

The first fold is used to extract the latent feature by the encoder;
The second fold is used as the pseudo ground-truth to compute the completion loss with the predicted shape;
The third fold is used to compute the latent reconstruction loss by encoding it and matching its latent feature to the first one?

Please let me know if I understood your design correctly.

[CuiRuikai]

1. Do you mean that the reconstruction loss is computed by using the predicted point cloud to fit a certain local region of the input point cloud (rather than fitting the entire complete shape)? Yes, this is described in the paper.
2. Yes.

The first fold is used to extract the latent feature by the encoder;
We use this fold to compute the reconstruction loss

[Frog-2915]

Thank you so much for your kind and patient explanations!
I have learned a lot from your detailed replies.
I truly appreciate your help and the time you spent answering my questions.

[CuiRuikai]

Thank you so much for your kind and patient explanations! I have learned a lot from your detailed replies. I truly appreciate your help and the time you spent answering my questions.

You are welcome. Let me know if you have further questions.