XRNeRF

Introduction

English | 简体中文

XRNeRF is an open-source PyTorch-based codebase for Neural Radiance Field (NeRF). It is a part of the OpenXRLab project.

https://user-images.githubusercontent.com/24294293/187131048-5977c929-e136-4328-ad1f-7da8e7a566ff.mp4

This page provides basic tutorials about the usage of XRNeRF. For installation instructions, please see installation.md.

• XRNeRF
  • Introduction
  • Benchmark
  • Datasets
  • Installation
  • Build a Model
    • Basic Concepts
    • Write a new network
  • Train a Model
    • Iteration Controls
    • Train
    • Test
  • Tutorials
  • Other Documents
  • Citation
  • License
  • Contributing
  • Acknowledgement
  • Projects in OpenXRLab

Benchmark

More details can be found in benchmark.md.

Supported scene-NeRF methods:

Datasets

It is recommended to symlink the dataset root to $PROJECT/data. If your folder structure is different, you may need to change the corresponding paths in config files.

xrnerf
├── xrnerf
├── docs
├── configs
├── test
├── extensions
├── data
│   ├── nerf_llff_data
│   ├── nerf_synthetic
│   ├── multiscale
│   ├── multiscale_google
│   ├── ...

For more information on data preparation, please see dataset_preparation.md

Installation

We provide detailed installation tutorial for XRNeRF, users can install from scratch or use provided dockerfile.

It is recommended to start by creating a docker image:

docker build -f ./docker/Dockerfile --rm -t xrnerf .

For more information, please follow our installation tutorial.

Build a Model

Basic Concepts

In XRNeRF, model components are basically categorized as 4 types.

• network: the whole nerf model pipeline, usually contains a embedder, mlp and render.
• embedder: convert point-position and viewdirection data into embedded data, embedder can be function only or with trainable paramters.
• mlp: use the output of embedder as input, and output raw data (the rgb and density value at sampled position) for render, usually contains FC layers.
• render: receive mlp's raw data, output the rgb value at a pixel.

Following some basic pipelines (e.g., NerfNetwork), the model structure can be customized through config files with no pains.

Write a new network

To write a new nerf network, you need to inherit from BaseNerfNetwork, which defines the following abstract methods.

• train_step(): forward method of the training mode.
• val_step(): forward method of the testing mode.

NerfNetwork is a good example which show how to do that.

To be specific, if we want to implement some new components, there are several things to do.

1. create a new file in xrnerf/models/networks/my_networks.py.

   from ..builder import NETWORKS
   from .nerf import NerfNetwork
   
   @NETWORKS.register_module()
   class MyNerfNetwork(NerfNetwork):
   
       def __init__(self, cfg, mlp=None, mlp_fine=None, render=None):
           super().__init__(cfg, mlp, mlp_fine, render)
   
       def forward(self, data):
           ....
   
       def train_step(self, data, optimizer, **kwargs):
           ....
   
       def val_step(self, data, optimizer=None, **kwargs):
           ....

2. Import the module in xrnerf/models/networks/__init__.py

   from .my_networks import MyNerfNetwork

3. modify the config file from

   model = dict(
       type='NerfNetwork',
       ....

   to

   model = dict(
       type='MyNerfNetwork',
       ....

To implement some new components for embedder/mlp/render, procedure is similar to above.

• To write a new nerf embedder, you need to inherit from nn.Module or BaseEmbedder, and define the forward method. BaseEmbedder is a good example.
• To write a new nerf mlp, you need to inherit from nn.Module or BaseMLP, and define the forward method. NerfMLP is a good example.
• To write a new nerf render, you need to inherit from nn.Module or BaseRender, and define the forward method. NerfRender is a good example.

Train a Model

Iteration Controls

XRNeRF use mmcv.runner.IterBasedRunner to control training, and mmcv.runner.EpochBasedRunner to for test mode.

In training mode, the max_iters in config file decide how many iters. In test mode, max_iters is forced to change to 1, which represents only 1 epoch to test.

Train

python run_nerf.py --config configs/nerf/nerf_blender_base01.py --dataname lego

Arguments are:

• --config: config file path.
• --dataname: select which data under dataset directory.

Test

We have provided model iter_200000.pth for test, download from here

python run_nerf.py --config configs/nerf/nerf_blender_base01.py --dataname lego --test_only --load_from iter_200000.pth

Arguments are:

• --config: config file path.
• --dataname: select which data under dataset directory.
• --test_only: influence on whole testset once.
• --load_from: load which checkpoint to test, this will overwrite the original load_from in config file to for convenience.

Tutorials

Currently, we provide some tutorials for users to

• learn about configs
• customize data pipelines
• model definition

Other Documents

Except for that，The document also includes the following

• api
• dataset
• installation
• benchmark
• FAQ

Citation

If you find this project useful in your research, please consider cite:

@misc{xrnerf,
    title={OpenXRLab Neural Radiance Field Toolbox and Benchmark},
    author={XRNeRF Contributors},
    howpublished = {\url{https://github.com/openxrlab/xrnerf}},
    year={2022}
}

License

The license of our codebase is Apache-2.0. Note that this license only applies to code in our library, the dependencies of which are separate and individually licensed. We would like to pay tribute to open-source implementations to which we rely on. Please be aware that using the content of dependencies may affect the license of our codebase. Some supported methods may carry additional licenses.

Contributing

We appreciate all contributions to improve XRNeRF. Please refer to CONTRIBUTING.md for the contributing guideline.

Acknowledgement

XRNeRF is an open source project that is contributed by researchers and engineers from both the academia and the industry. We appreciate all the contributors who implement their methods or add new features, as well as users who give valuable feedbacks. We wish that the framework and benchmark could serve the growing research community by providing a flexible framework to reimplement existing methods and develop their own new models.

Projects in OpenXRLab

• XRPrimer: OpenXRLab foundational library for XR-related algorithms.
• XRSLAM: OpenXRLab Visual-inertial SLAM Toolbox and Benchmark.
• XRSfM: OpenXRLab Structure-from-Motion Toolbox and Benchmark.
• XRLocalization: OpenXRLab Visual Localization Toolbox and Server.
• XRMoCap: OpenXRLab Multi-view Motion Capture Toolbox and Benchmark.
• XRMoGen: OpenXRLab Human Motion Generation Toolbox and Benchmark.
• XRNeRF: OpenXRLab Neural Radiance Field (NeRF) Toolbox and Benchmark.