search

nerfstudio-project / nerfstudio

A collaboration friendly studio for NeRFs
https://docs.nerf.studio
Apache License 2.0
9.35k stars 1.26k forks source link

High resolution data training #2045

Open Yaodizhao opened 1 year ago

Yaodizhao commented 1 year ago

Hello, thank you for your outstanding work. May I ask what is the reason for the problem and how to modify it? When using 8120 * 5460 high-resolution photos for nerfacto training, four 80G GPUs are used for training. However, each time there is an issue similar to "process 2 terminated with signal SIGKILL"

Yaodizhao commented 1 year ago

I feel that the problem may be due to caching all images. What is the reason for this

tancik commented 1 year ago

You don't really get much benefit with images >2k, I'd recommend decreasing your image size.

Yaodizhao commented 1 year ago

What is the reason for this? Is it because sampling only selects 2K rays for sampling? If high-resolution images are available, can we choose to adjust the parameters and select more rays for training, resulting in higher resolution of the results

f-dy commented 1 year ago

This is because of the limited resolution of the NeRF itself. What you could do, if you really want to play with higher resolution, is use less images for training, and shuffle the training images. For examples, with 700 images, --pipeline.datamanager.train-num-images-to-sample-from=200 --num-times-to-repeat-images=1000 will pick 200 new images from the 700 available every 1000 iterations. Sample (impressive) result here on discord