Problem when using large scale about ngp_pl HOT 9 CLOSED

When digging in a bit more I found the model is actually evaluating a lot more samples per ray with the new code (often the MAX_SAMPLE) was the old calc_dt returning a too low number of samples to evaluate?

from ngp_pl.

I'm still debugging to match the original implementation for real scenes (scale>0.5). Therefore anything with scale>0.5 was never tested and never used...

from ngp_pl.

Hi, just push the code to support large scale now. Please tell me if the problem persists

from ngp_pl.

I see thanks, reducing the batch_size or the MAX_SAMPLES solves the problem for now so it was just a memory problem throwing an uninformative cuda error.. Also I'm using an RTX A6000 so probably the problem is more severe with other GPUs. Thanks!

from ngp_pl.

Hi, just push the code to support large scale now. Please tell me if the problem persists

Great, thanks will try now!

from ngp_pl.

Unfortunately, it seems like it's not solved, I still have to make the batch_size smaller. Also now where it should predict objects with large depths it puts objects very close to the camera. And when training instant-ngp with the same data and settings this doesn't happen.

from ngp_pl.

What scene are you using that require such a large scale? First if the scale is large, it means we need more density grids, so more memory consumption, therefore we can only feed a smaller batch size, this is reasonable.
Second, "it puts objects very close to the camera" this problem actually occurs in one of my test data, and I'm still debugging for it, but others (like the fox data and other llff_data) work fine for now. If you have any idea you're welcome to share!

from ngp_pl.

Yes I agree memory consumption makes sense, I was just surprised that from one version to a few later there was a large increase, but I guess the first version I used wasn't sampling properly along rays for large scale factors and the new one is. I'm working on driving scenes, I'll try to debug that too on my scene and let you know what I find, thanks!

from ngp_pl.

The phenomenon of "putting everything in front of each camera" still occurs on some scenes, and I also observe this on the original instant-ngp (some scenes cannot train successfully, or trains roughly good but with a lot of floaters). If someone has insight on it, you are welcome to share.

That being said, among many data I've tried (nerf_llff_data, mipnerf 360 data and fox, with different scales varying from 2.0 to 4.0), most of them train successfully, so I think the scale problem is resolved.

from ngp_pl.