A Theoratical Question about CPN

[JUGGHM]

Hi, Yanchao!
You mentioned in the appendix of your paper that when ψ imposes no compression on I, training the exponential loss leads to an estimation of P(f|I). But I wonder why it is compression. In other words, why you part the network to an encoder and a decoder? I would be grateful if you can share some knowledge about the proposal of CPN. I like the probability model very much. It is very logical.

[YanchaoYang]

Thanks for your interest!
I am not sure what you mean by part the network to an encoder and a decoder.
But the idea of using encoder-decoder is to impose an information bottleneck, such that the function Q in Eq. 3 is informative. The skip connection makes sure that the image I is not compressed, in order to make Q maximumly informative. As derived in the Appendix, if I is compressed, Q will approximate some induced probability instead of P(f|I).

[JUGGHM]

Thanks for your explanation!!!! I think I should read some materials about information bottleneck to fully understand the structural design of CPN.

I have an additional question about your work DDP(in CVPR2019). Your mentioned that the stereo supervision item loss leads to RMSE performance in unsupervised training (Of Course). But we know that in KITTI test that there are no stereo pairs provided. In Table.2, you demonstrate that the stereo loss item of Eq.(9) can improve the performance.

Then how the stereo loss items are computed in KITTI test set?

Looking forward to your reply and guidance!