search

bennyguo / instant-nsr-pl

Neural Surface reconstruction based on Instant-NGP. Efficient and customizable boilerplate for your research projects. Train NeuS in 10min!
MIT License
857 stars 84 forks source link

Results for training on DTU #111

Closed fatbao55 closed 1 year ago

fatbao55 commented 1 year ago

Thanks for the great work!

I would like to check what is your PSNR for training on DTU. Mine is 3.137 and 6.637 without and with mask respectively. It seems rather low compared to the values reported when trained on Nerf-Synthetic (30+).

Is this abnormal?

AIBluefisher commented 1 year ago

It is absolutely abnormal. The PSNR on DTU should all be larger than 20.

anonymous-pusher commented 1 year ago

You might have skipped the part in the readme saying that :

PSNR in the testing stage is meaningless, as we simply compare to pure white images in testing.

The current dataloader for dtu does not load ground truth images, but only uses white pixels. You will need to define your own train/test splits in the dataloader or set it to use the training views during testing. you can check the line https://github.com/bennyguo/instant-nsr-pl/blob/e5fe3f246cf2d512494a73c727174c3ca1c3c695/datasets/dtu.py#L124

fatbao55 commented 1 year ago

@mJones00 You're right, thanks for the clarification!