Hello everyone, thank you for sharing this experiments for this wonderfull project! Which version of Jax did you use? and could you share the requirements? I am trying to replicate the experiments and, when I download the Jax and jaxlibrary as mentioned on the readme I get this error :/

Thanks for releasing the JAX based code! I am trying to use your NeRF implementation as a base to implement DietNeRF (in JAX and Haiku). But I noticed your implementation has 2 differences with the original one, as stated in your paper:

In our view synthesis experiments, we use a simplified NeRF model (simple-NeRF) that maintains the same image supervision and volume render-ing context. Unlike the original NeRF model, we do not feed in the viewing direction and we use a single model in-stead of the two “coarse” and “fine” models used by NeRF.

Would you mind telling me the motivation behind this simplification?
(e.g. Is it because the original NeRF didnt work in your setting? or because its easier to implement this in JAX?) Knowing this may help me to debug my DietNeRF implementation in the future. Many thanks!

in function get_example, you have:

    if 'train' in split:
        img_idx + 25
    if 'val' in split:
        img_idx + 20

assignment operations are missing. it should be:

    if 'train' in split:
        img_idx += 25
    if 'val' in split:
        img_idx += 20

The phototourism experiment sets the final dists vaule to a very big number, like 1e10: dists = np.concatenate([z_vals[..., 1:] - z_vals[..., :-1], np.broadcast_to([1e10], z_vals[...,:1].shape)], -1)

On the other hand, the shapenet experiment uses a small number 1e-3 for the final dists: dists = np.concatenate([z_vals[..., 1:] - z_vals[..., :-1], np.broadcast_to([1e-3], z_vals[...,:1].shape)], -1)

What does it mean to have a very small or large value for the final dists? As I understand, the scene contents are expected to be bounded between near and far plane. So does it matter what value I set for the final dists?

Sorry if this is a silly question. I couldn't find an answer after searching through other nerf repos.

Hi, thanks for the great work! It is really cool to see how simple meta-learning algorithms can be useful for phototourism.

But I couldn't find any splits file mentioning which images to use for test and validation. How do you make the split for different landmark scenes?

I would like to ask about your experimental part in Single image view synthesis for ShapeNet, you are using only one image in the meta-learning stage under the setting of SV Meta, which means that only one image is used for the whole training process to generate the class(car or chair) 360 degree reconstruction effect for all instances?

I want to know how some parameters, such as the camera's internal parameters, external parameters, etc., are obtained when processing shapenet datasets

I have made a PyTorch implementation of Photo-Tourism and ShapeNet experiments from the paper.

Please check it out here: https://github.com/sanowar-raihan/nerf-meta

I created an accessible colab notebook for fitting SIREN on image/video + option to load your learned initializations for image regression, including using them to initialize video fitting. please find here: colab.research.google.com/gist/eyaler/20a9037a3619378b276b2303dadb558d