tf-garf

An unofficial implementation of Gaussian Activated Radiance Fields in TensorFlow 2.

Original paper by Chng et al. (2022): [arXiv] [webpage]

See the notebook for more info.

Now that the official code is available in PyTorch, you'll probably want to use that code instead!

Data

This implementation expects LLFF-style, forward-facing data, such as the NeRF LLFF dataset.

from lib.data import load_data
data = load_data('path/to/data/root')

If you have other data, you can use it like so.

from lib.data import hwf_rgb_to_data

hwf = np.array([image_height, image_width, focal_length]) # in pixels
rgb = # ... list of np.array objects of shape (image_height, image_width, 3)

data = hwf_rgb_to_data(hwf, rgb)

Model

Once you have the data, you can create a model object.

from model.garf import GaussianRadianceField
model = GaussianRadianceField(data, num_samples=128)

You can load the pre-trained model for the NeRF flower dataset. Pass opt=True to load the optimizer state.

model.load('pretrain/flowers', opt=False)

Render a view from a specific image index or pose with model.garf.GaussianRadianceField.render.