Using custom data

Training model on existing datasets is only so fun. If you would like to train on self captured data you will need to process the data into an existing format. Specifically we need to know the camera poses for each image. COLMAP is a standard tool for extracting poses. It is possible to use other methods like SLAM or hardware recorded poses. We intend to add documentation for these other methods in the future.

Nerfactory dataset

To assist running on custom data we have a script that will process a video or folder of images into a format that is compatible with nerfstudio. We use COLMAP and FFmpeg in our data processing script, please have these installed. We have provided a quickstart to installing COLMAP below, FFmpeg can be downloaded from here

To process your data run:

python scripts/process_data.py --data FOLDER_OR_VIDEO --output-dir PROCESSED_DATA_DIR

If your data was captured with a fisheye camera add --camera-type fisheye.

:::{admonition} Tip :class: info

COLMAP can be finicky. Try your best to capture overlapping, non-blurry images. :::

Training on your data

Simply specify that you are using the nerfactory dataparser and point the data directory to your processed data.

python scripts/train.py instant-ngp nerfactory --pipeline.datamanager.dataparser.data-directory PROCESSED_DATA_DIR

Installing COLMAP

The COLMAP maintainers recommend using VCPKG to install COLAMP. If the following command do not work, please refer to the COLMAP installation guide for additional installation methods.

::::::{tab-set} :::::{tab-item} Linux

::::{tab-set} :::{tab-item} CUDA

git clone https://github.com/microsoft/vcpkg
cd vcpkg
./bootstrap-vcpkg.sh
./vcpkg install colmap[cuda]:x64-linux

::: :::{tab-item} CPU

git clone https://github.com/microsoft/vcpkg
cd vcpkg
./bootstrap-vcpkg.sh
./vcpkg install colmap:x64-linux

::: ::::

:::::

:::::{tab-item} OSX

git clone https://github.com/microsoft/vcpkg
cd vcpkg
./bootstrap-vcpkg.sh
./vcpkg install colmap

:::::

:::::{tab-item} Windows