dawning77/nerrf

NeRRF: 3D Reconstruction and View Synthesis for Transparent and Specular Objects with Neural Refractive-Reflective Fields

NeRRF

---

NeRRF: 3D Reconstruction and View Synthesis for Transparent and Specular Objects with Neural Refractive-Reflective Fields

Setup

conda env create -f environment.yml
conda activate nerrf

Dataset

Our blender synthetic dataset can be found here: blender_dataset

Usage

1. In the NeRRF directory, ensure you have the following data:

   NeRRF
   |-- data
       |-- blender
           |-- depth
           |-- meta
           |-- specular
           |-- transparent
   

2. Run the training script

   # reconstruct the specular horse of the blender dataset
   sh scripts/train_horse_s.sh
   

   The scripts can be switched from the geometry reconstruction stage to the radiance estimating stage by changing the value of the stage variable. Set stage to 1 for geometry reconstruction.

3. Run the evaluation script

   # evaluate the specular horse of the blender dataset
   sh scripts/train_horse_s.sh
   

nerrf's Issues

About the loading data when implemetation

Hi, Thanks for the great works

When I was trying to implement the project, I encountered an issue with the data of the Nerf file. An error like this

Then I went to the file; it seems the code gets the vertices from the npz file based on the grid size. But I didn't find any data about this. May I know where these files are so I can confirm if there is a problem with my environment installation?

训练完结果如何显示

请问训练完的结果该如何显示出来呢？

The environment.yml seems to have some version conflicts

For example, kaolin 0.12.0 is not compatible with scipy 1.9.3.

kaolin 0.12.0 requires scipy<=1.7.2,>=1.2.0, but you have scipy 1.7.3 which is incompatible.

And I installed the environment from scratch without torch in the environment. When the script installed chamferdist==1.0.0, it ran into an import error because the setup.py for chamferdist needed to import torch, which did not exist in my conda environment by then. So maybe you can move the torch requirement to the top of the pip requirement list?

Running on NeRF Synthetic Dataset

Dear Authors,

Thank you so much for the brilliant work and for sharing the code!

May I ask if it is possible to run the code on NeRF synthetic dataset? Specifically, when loading the camera poses "cam_world_pose" from meta jsons, does the code expect the cam_world_pose to be in Blender format, or openGL format? What is the difference between "cam_world_pose" and "transform_matrix" in NeRF synthetic dataset?

Many thanks!