search

liuyuan-pal / SyncDreamer

[ICLR 2024 Spotlight] SyncDreamer: Generating Multiview-consistent Images from a Single-view Image
https://liuyuan-pal.github.io/SyncDreamer/
MIT License
906 stars 39 forks source link

About NVS metrics calculation #46

Closed PikaCat-OuO closed 1 year ago

PikaCat-OuO commented 1 year ago

I have a simple question that I still haven't figured out so far. When calculating metrics like PSNR, do you directly use the image output from the model to compare with the ground truth GSO images, or first go through the reconstruction process and render images from the reconstructed object and then compare with the ground truth?

liuyuan-pal commented 1 year ago

Hi, I use the generated views to calculate the PSNR instead of the rendering views.

PikaCat-OuO commented 1 year ago

Thanks, get it.😊

PikaCat-OuO commented 1 year ago

btw, regarding the input images, when performing generation, do we directly use the first image from ground truth i.e. "eval_samples/chicken-gt/000.jpg" as inputs?

liuyuan-pal commented 1 year ago

Yes, we use the first image as condition.