Regarding to training weights

[batman47steam]

Hi, I retrained the model and found that within 50 training epochs, the performance metrics on the simulated dataset kept improving, even surpassing those of the publicly released pretrained weights. However, the reconstruction quality on real images deteriorated, showing more noticeable noise. I would like to ask for advice on how to select the final model weights for testing.

[Depth2World]

This phenomenon does exist. The dilemma is the lack of a real-world ground truth for evaluation. My suggestion is to select the most challenging scenario when validating in the real world and ultimately choose the one with the best realistic effect.

[batman47steam]

Got it ! Thanks for the advice !

[batman47steam]

I tested the traditional methods such as FK using synthetic data, and while the reconstructed intensity images look reasonable in most cases, the corresponding depth maps appear almost meaningless.
Is this expected? I noticed that in the paper, the traditional method seems to produce some reasonably depth reconstructions.