How much is the final training loss? about cvp-mvsnet HOT 3 CLOSED

Hi Todd-Qi,

Thank you for your interest in our project.

Regarding your questions:
The final training loss of our pre-trained model is also around 50, so your final loss is reasonable.
However we used batch size 16 on training. Smaller batch size may cause more noise during training.
And about your result, what numbers have you got on DTU dataset?
Could you please provide the evaluation parameters and fusion parameters you used to get the result?

Cheers,
Jiayu

from cvp-mvsnet.

Hello, @JiayuYANG
I'm using the same fusion parameter, i.e. prob_threshold=0.8, disp_threshold=0.13 and num_consistent=3.
I think the problem comes from the predicted depth map. The depth map of mine has many holes.



from the top to bottom, (a)the original ref image, (b)the depth map predicted by my trained model, (c)the depth map predicted by your pretrained model.
For the evaluation, the number of depth sampling is 48.

In addition, I compute the mean absolute error(MAE) on the DTU validtion set. The MAE of your pretrained model is about 10 while my trained model is about 15.

Do you have any suggestions how to solve this problem?
Thank You！

from cvp-mvsnet.

Hi Todd-Qi,

I have updated our code to solve this issue,
you may pull the latest version to reproduce our results.

Training from scratch using default settings, the latest version can produce following results on my machine.

Cheers,
Jiayu

from cvp-mvsnet.