This repo enables active data collection to add during training, built off the amazing work done by the nerfstudio team, and is the codebase for Data Augmentation for NeRFs in the Low Data Limit [submitted to ICRA '25].

Quickstart

Install

Follow the installation procedure from Nerfstudio; our code was built off commit 516fd7c. We run our code on a server with Ubuntu 20.04 and an Nvidia RTX 6000 Ada card.

To summarize:

git clone git@github.com:Agaggar/nerfstudio-low-data.git
cd nerfstudio
pip install --upgrade pip setuptools
pip install -e .

Data

You'll also need some data to train your NeRF model with. We provide sample files (from the lego dozer, scaled down 0.1x) in data/lego_scale

The easiest way to get sample data is to download the Blender dataset from the original NeRF paper (nerf_synthetic.zip). You can then follow the Nerfstudio guide to training with Blender data here.

Alternatively, you can record a video in real life and use COLMAP to label poses with images (following this guide by Nvidia's amazing instant NGP team).

Blender

If you want to augment additional images in Blender, download Blender software; we run Blender v4.1.1 on Linux. The blend files we use are located data/blender_files, and are scaled down by 0.1x.

Using our code

Running ns-train nerfacto-active --data PATH_TO_DATA --random-seed POSITIVE_INTEGER sets the default configuration for active, iterative training.

For example:

ns-train nerfacto-active --data data/lego_scale --random-seed 0 --subset-data 10 --max-data 15 --add-amount 5 --data-selector nbb

Additional flags:

• --iterative_training: bool = False: Whether to train iteratively (add data during training)
• --steps_per_iterative_add: Number of steps between adding data. i.e., 200 means add add_amount data every 200 iterations until max-data reached.
• --subset_data: How many images to start training with
• --random_seed: Random seed for sampling data from total training set. Must be specified for data augmenting.
• --max_data: max number of data to add to in training set
• --add_amount: how many images to add between each training step
• --data_selector: if iterative training, method for adding data. currently supports uniform and our method (nbb)

Citing

If you find this code useful for your research or the use data generated by our method, please consider citing our (after release).