search

alievk / npbg

Neural Point-Based Graphics
MIT License
324 stars 51 forks source link

The viewer's results were not as good as the training #26

Closed JOP-Lee closed 2 years ago

JOP-Lee commented 2 years ago

Hi, thanks for the good work. When I trained the scene, the PSNR was up to 25, which was clear to see on the Tensorboard. Why is it not as good as the one seen in the viewer during training, and the PSNR obtained is less than 20? (the same pose and point cloud) Does the viewer have any tricks to save the image and reduce distortion?

seva100 commented 2 years ago

Hi @JOP-Lee! Usually we have not really experienced a lot of overfitting with Neural Point-Based Graphics, so it's a bit strange that this happens. May I ask which scene do you train it for and maybe you have some screenshots and details to confirm the effect?

JOP-Lee commented 2 years ago

Hi @JOP-Lee! Usually, we have not really experienced a lot of overfitting with Neural Point-Based Graphics, so it's a bit strange that this happens. May I ask which scene do you train it for and maybe you have some screenshots and details to confirm the effect?

@seva100 hi, Seva, I am training a beautiful scene like this. I got the problem when I train and test the same pose and point cloud. if there has overfitting, I think the training data set result won't be worse, right? 195

seva100 commented 2 years ago

I think I will need more details to understand it. Can you please share your scene, the trained checkpoint and the train.py & viewer.py launch commands?

JOP-Lee commented 2 years ago

@seva100 thank you. I have found the problem. The reason is that I made a mistake in the output image. Thank you for your reply.