search

Kai-46 / nerfplusplus

improves over nerf in 360 capture of unbounded scenes
BSD 2-Clause "Simplified" License
911 stars 101 forks source link

Model output for demo examples are blurry #30

Open zhanibekrysbek opened 3 years ago

zhanibekrysbek commented 3 years ago

After training with default configs listed in repo, I got much blurrier rendered images compared to what has been demonstrated. In particular truck scene has been trained for 500K iterations. Same situation for the train scene. I might be missing something important in the training phase.

What are the training params for the best model performance with high resolutions?

For reference see below two images, one is rendered second is the ground truth.

image

image

Kai-46 commented 3 years ago

Did you use the provided images and provided configuration parameters?

zhanibekrysbek commented 3 years ago

Yes. I used the default settings on the truck and train data set.

Kai-46 commented 3 years ago

Okay. How many GPUs were used for parallel training then?

zhanibekrysbek commented 3 years ago

One V100 gpu with 32Gb. Model apparently uses up to 4Gb..

Kai-46 commented 3 years ago

Oh, you might want to check this line of code. I determine the number of rays to sample according to the available GPU memory at each training iteration; but I only take into account GPU with 12G or 24G mem. I used 4 12G-GPU during training, which means 2048 rays are sampled each iteration. In your case, can you try increasing the number of rays sampled at each iteration to 2048?

zhanibekrysbek commented 3 years ago

Got it! What param values (N_rand, chunk_size) you suggest for my settings for example?

Kai-46 commented 3 years ago

Since your GPU has 32G mem, it might support a maximum N_rand=1300. chunk_size=10900. But if you only have access to this single GPU, it might be easier for you just to double the number of training iterations, as you are using half the number of rays I used.