NeRF

NeRF Docker Setup

Prerequisites

  1. Docker

  2. Latest nvidia drivers

  3. CUDA 11.8 capable GPU

Install

  1. Pull latest docker image

docker pull dromni/nerfstudio:<version_number>

You can see the latest tag in this link https://hub.docker.com/r/dromni/nerfstudio/tags

# example command
docker pull dromni/nerfstudio:0.1.19
  1. Run docker image

docker run --gpus all -v /folder/of/your/data:/workspace/ -v /home/<YOUR_USER>/cache/:/home/user/.cache/ -p 7007:7007 --rm -it dromni/nerfstudio:0.1.13

Example for Windows:

docker run --gpus all -v D:\nerf:/workspace/ -v D:\nerf/:/home/user/.cache/ -p 7007:7007 --rm -it dromni/nerfstudio:0.1.19
-  Give the container access to nvidia GPU (required).
-  Mount a folder from the local machine into the container to be able to process them (required).
-  Mount cache folder to avoid re-downloading of models everytime (recommended).
-  Map port from local machine to docker container (required to access the web interface UI).
-  Remove container after it is closed (recommended).
-  Start container in interactive mode.
-  Docker image name

Training First Model

Don't forget to change the directory

cd /workspace

Be sure to do everything you do after this point in the /workspace directory.

Start Training

ns-download-data nerfstudio --capture-name=poster

Because this code is often used too much, google drive can put a download restriction, read the error message and try to download the file in the link with your browser.

There are also different dataset you can use, to show them run this code

ns-download-data nerfstudio --help
bww_entrance, campanile, desolation, library, 
poster, redwoods2, storefront, vegetation, Egypt, 
person, kitchen, plane, dozer, floating-tree, aspen, 
stump, sculpture, Giannini-Hall, all, nerfstudio-dataset

After select one of the dataset you can download as well with following command

ns-download-data nerfstudio --capture-name=bmw_entrance

You will see all the downloaded datasets in this directory D:\nerf\data\nerfstudio

Explore NeRF Studio

Start training with the following command, you will need change the datapath to your dataset folder

ns-train nerfacto --data data/nerfstudio/poster

or

ns-train nerfacto --data data/nerfstudio/bww_entrance

If everything is correct, ideal output would be like this.

A web page similar to the link below will appear on your terminal screen, open this page and continue.

https://viewer.nerf.studio/versions/22-12-02-0/?websocket_url=ws://localhost:7007 

Resume & Stop Training

You can stop the training with the Ctrl + c shortcut.

If you want to continue the training from where it left off, run the code below.

ns-train nerfacto --data data/nerfstudio/poster --load-dir {outputs/.../nerfstudio_models}

Example code

ns-train nerfacto --data data/nerfstudio/bww_entrance/ --load-dir outputs/bww_entrance/nerfacto/2023-03-14_183456/nerfstudio_models/

Visualize Only (without training)

ns-viewer --load-config {outputs/.../config.yml}
# example code
ns-viewer --load-config outputs/bww_entrance/nerfacto/2023-03-14_183456/config.yml

Training on Custom Data

Put your files in the main directory of the volume you opened, you can also keep them in a folder in the main directory.

ns-process-data {video,images,polycam,insta360,record3d} --data {DATA_PATH} --output-dir {PROCESSED_DATA_DIR}
# example code
ns-process-data video --data raw_video/forest_walk.mp4 --output-dir forest_walk_processed

Arguments:

ns-process-data images [-h] --data PATH --output-dir PATH
                       [--camera-type {perspective,fisheye}]
                       [--matching-method {exhaustive,sequential,vocab_tree}]
                       [--sfm-tool {any,colmap,hloc}]
                       [--feature-type 
{any,sift,superpoint,superpoint_aachen,superpoint_max,superpoint_inloc,r2d2,d2
net-ss,sosnet,disk}]
                       [--matcher-type 
{any,NN,superglue,superglue-fast,NN-superpoint,NN-ratio,NN-mutual,adalam}]
                       [--num-downscales INT] [--skip-colmap]
                       [--colmap-cmd STR] [--no-gpu] [--verbose]
ns-train nerfacto --data {PROCESSED_DATA_DIR}
# example code
ns-train nerfacto --data forest_walk_processed/

Example

It points to the same place in the following two paths. one its location on the host and the other its location inside the container.

Raw File Path:

  • D:\nerf\raw_data\IMG_6070.MOV

  • /workspace/raw_data/IMG_6070.MOV

Target Folder Path:

  • D:\nerf\outputs\test_video

  • /workspace/outputs/bank

Preparing

ns-process-data video --data /workspace/raw_data/IMG_6070.MOV --output-dir /workspace/outputs/bank

Training

ns-train nerfacto --data outputs/bank

KIRI Engine

Coming soon...

Last updated