search

google / mipnerf

Apache License 2.0
894 stars 109 forks source link

Fails at high resolution on LLFF dataset #15

Open cwchenwang opened 2 years ago

cwchenwang commented 2 years ago

Hi, I appreciate your excellent work. I found that when training mip-nerf on LLFF scene (fern specifically), it reconstructs well on original (252x189) and lower resolution. color_000

However, when I try to render higher resolution images (e.g. 512x378 here), they contain a lot of noise that vanilla nerf doesn't contain. What might be the possible reason? color_001

jonbarron commented 2 years ago

It sounds like you're training at a low resolution, and then testing at a higher resolution than you trained at, is that correct? If that is what you're doing, then these artifacts make sense. There's nothing in mip-NeRf that should enable super-resolution or generalization to higher resolutions than what you trained at. All it should be able to do is interpolate between the scales you have trained with. Vanilla NeRF might behave better in this circumstance as it has no understanding of scale.