This project attempted to implement the paper Putting NeRF on a Diet (DietNeRF) in JAX/Flax.
DietNeRF is designed for rendering quality novel views in few-shot learning scheme, a task that vanilla NeRF (Neural Radiance Field) struggles.
To achieve this, the author coins Semantic Consistency Loss to supervise DietNeRF by prior knowledge from CLIP Vision Transformer. Such supervision enables DietNeRF to learn 3D scene reconstruction with CLIP's prior knowledge on 2D views.
Besides this repo, you can check our write-up and demo here:
Or you can set up our Streamlit demo locally (model checkpoints will be fetched automatically upon startup)
shell
1pip install -r requirements_demo.txt
2streamlit run app.py
Streamlit Demo
✨ Implementation
Our code is written in JAX/ Flax and mainly based upon jaxnerf from Google Research. The base code is highly optimized in GPU & TPU. For semantic consistency loss, we utilize pretrained CLIP Vision Transformer from transformers library.
To learn more about DietNeRF, our experiments and implementation, you are highly recommended to check out our very detailed Notion write-up!
1# Clone the repo2git clone https://github.com/codestella/putting-nerf-on-a-diet
3# Create a conda environment, note you can use python 3.6-3.8 as4# one of the dependencies (TensorFlow) hasn't supported python 3.9 yet.5conda create --name jaxnerf python=3.6.12; conda activate jaxnerf
6# Prepare pip7conda install pip; pip install --upgrade pip
8# Install requirements9pip install -r requirements.txt
10# [Optional] Install GPU and TPU support for Jax11# Remember to change cuda101 to your CUDA version, e.g. cuda110 for CUDA 11.0.12!pip install --upgrade jax "jax[cuda110]" -f https://storage.googleapis.com/jax-releases/jax_releases.html
13# install flax and flax-transformer14pip install flax transformers[flax]
⚽ Dataset
Download the datasets from the NeRF official Google Drive.
Please download the nerf_synthetic.zip and unzip them
in the place you like. Let's assume they are placed under /tmp/jaxnerf/data/.
💖 Methods
👉👉 You can check VEEEERY detailed explanation about our project on Notion Report
스크린샷 2021-07-04 오후 4 11 51
Based on the principle
that “a bulldozer is a bulldozer from any perspective”, Our proposed DietNeRF supervises the radiance field from arbitrary poses
(DietNeRF cameras). This is possible because we compute a semantic consistency loss in a feature space capturing high-level
scene attributes, not in pixel space. We extract semantic representations of renderings using the CLIP Vision Transformer, then
maximize similarity with representations of ground-truth views. In
effect, we use prior knowledge about scene semantics learned by
single-view 2D image encoders to constrain a 3D representation.
You can check detail information on the author's paper. Also, you can check the CLIP based semantic loss structure on the following image.
스크린샷 2021-07-04 오후 4 11 51
Our code used JAX/FLAX framework for implementation. So that it can achieve much speed up than other NeRF codes. At last, our code used hugging face, transformer, CLIP model library.
You can toggle the semantic loss by “use_semantic_loss” in configuration files.
💎 Experimental Results
❗ Rendered Rendering images by 8-shot learned Diet-NeRF
DietNeRF has a strong capacity to generalise on novel and challenging views with EXTREMELY SMALL TRAINING SAMPLES!
HOTDOG / DRUM / SHIP / CHAIR / LEGO / MIC
❗ Rendered GIF by occluded 14-shot learned NeRF and Diet-NeRF
We made artificial occlusion on the right side of image (Only picked left side training poses).
The reconstruction quality can be compared with this experiment.
DietNeRF shows better quality than Original NeRF when It is occluded.
😎 What we improved from original JAX-NeRF : Innovation
Neural rendering with fewshot images
Hugging face CLIP based semantic loss loop
You can choose coarse mlp / coarse + fine mlp training
(coarse + fine is on the main branch / coarse is on the coarse_only branch)
coarse + fine : shows good geometric reconstruction
coarse : shows good PSNR/SSIM result
Make Video/GIF rendering result, --generate_gif_only arg can run fast rendering GIF.
Cleaning / refactoring the code
Made multiple models / colab / space for Nice demo
💞 Social Impact
Game Industry
Augmented Reality Industry
Virtual Reality Industry
Graphics Industry
Online shopping
Metaverse
Digital Twin
Mapping / SLAM
🌱 References
This project is based on “JAX-NeRF”.
@software{jaxnerf2020github,
author = {Boyang Deng and Jonathan T. Barron and Pratul P. Srinivasan},
title = {{JaxNeRF}: an efficient {JAX} implementation of {NeRF}},
url = {https://github.com/google-research/google-research/tree/master/jaxnerf},
version = {0.0},
year = {2020},
}
This project is based on “Putting NeRF on a Diet”.
@misc{jain2021putting,
title={Putting NeRF on a Diet: Semantically Consistent Few-Shot View Synthesis},
author={Ajay Jain and Matthew Tancik and Pieter Abbeel},
year={2021},
eprint={2104.00677},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
Thank you for our mentor Suraj and organizers in JAX/Flax Community Week!
Our team grows up with this community learning experience. It was wonderful time!
스크린샷 2021-07-04 오후 4 11 51
Common Computer AI sponsored multiple V100 GPUs for our project!
Thank you so much for your support!
