GRF: Learning a General Radiance Field for 3D Representation and Rendering

Related tags

Deep LearningGRF
Overview

GRF: Learning a General Radiance Field for 3D Representation and Rendering

[Paper] [Video]

GRF: Learning a General Radiance Field for 3D Representation and Rendering
Alex Trevithick1,2 and Bo Yang2,3
1Williams College, 2University of Oxford, 3The Hong Kong Polytechnic University in ICCV 2021

This is the codebase which is currently a work in progress.

Overview of GRF

GRF is a powerful implicit neural function that can represent and render arbitrarily complex 3D scenes in a single network only from 2D observations. GRF takes a set of posed 2D images as input, constructs an internal representation for each 3D point of the scene, and renders the corresponding appearance and geometry of any 3D point viewing from an arbitrary angle. The key to our approach is to explicitly integrate the principle of multi-view geometry to obtain features representative of an entire ray from a given viewpoint. Thus, in a single forward pass to render a scene from a novel view, GRF takes some views of that scene as input, computes per-pixel pose-aware features for each ray from the given viewpoints through the image plane at that pixel, and then uses those features to predict the volumetric density and rgb values of points in 3D space. Volumetric rendering is then applied.

Setting Up the Environment

Use conda to setup an environment as follows:

conda env create -f environment.yml
conda activate grf

Data

  • SRN cars and chairs datasets can be downloaded from the paper's drive link
  • NeRF-Synthetic and LLFF datasets can be downloaded from the NeRF drive link
  • MultiShapenet dataset can be downloaded from the DISN drive link

Training and Rendering from the Model

To train and render from the model, use the run.py script

python run.py --data_root [path to directory with dataset] ] \
    --expname [experiment name]
    --basedir [where to store ckpts and logs]
    --datadir [input data directory]
    --netdepth [layers in network]
    --netwidth [channels per layer]
    --netdepth_fine [layers in fine network]
    --netwidth_fine [channels per layer in fine network]
    --N_rand [batch size (number of random rays per gradient step)]
    --lrate [learning rate]
    --lrate_decay [exponential learning rate decay (in 1000s)]
    --chunk [number of rays processed in parallel, decrease if running out of memory]
    --netchunk [number of pts sent through network in parallel, decrease if running out of memory]
    --no_batching [only take random rays from 1 image at a time]
    --no_reload [do not reload weights from saved ckpt]
    --ft_path [specific weights npy file to reload for coarse network]
    --random_seed [fix random seed for repeatability]
    --precrop_iters [number of steps to train on central crops]
    --precrop_frac [fraction of img taken for central crops]
    --N_samples [number of coarse samples per ray]
    --N_importance [number of additional fine samples per ray]
    --perturb [set to 0. for no jitter, 1. for jitter]
    --use_viewdirs [use full 5D input instead of 3D]
    --i_embed [set 0 for default positional encoding, -1 for none]
    --multires [log2 of max freq for positional encoding (3D location)]
    --multires_views [log2 of max freq for positional encoding (2D direction)]
    --raw_noise_std [std dev of noise added to regularize sigma_a output, 1e0 recommended]
    --render_only [do not optimize, reload weights and render out render_poses path]
    --dataset_type [options: llff / blender / shapenet / multishapenet]
    --testskip [will load 1/N images from test/val sets, useful for large datasets like deepvoxels]
    --white_bkgd [set to render synthetic data on a white bkgd (always use for dvoxels)]
    --half_res [load blender synthetic data at 400x400 instead of 800x800]
    --no_ndc [do not use normalized device coordinates (set for non-forward facing scenes)]
    --lindisp [sampling linearly in disparity rather than depth]
    --spherify [set for spherical 360 scenes]
    --llffhold [will take every 1/N images as LLFF test set, paper uses 8]
    --i_print [frequency of console printout and metric loggin]
    --i_img [frequency of tensorboard image logging]
    --i_weights [frequency of weight ckpt saving]
    --i_testset [frequency of testset saving]
    --i_video [frequency of render_poses video saving]
    --attention_direction_multires [frequency of embedding for value]
    --attention_view_multires [frequency of embedding for direction]
    --training_recon [whether to render images from the test set or not during final evaluation]
    --use_quaternion [append input pose as quaternion to input to unet]
    --no_globl [don't use global vector in middle of unet]
    --no_render_pose [append render pose to input to unet]
    --use_attsets [use attsets, otherwise use slot attention]

In particular, note that to render and test from a trained model, set render_only to True in the config.

Configs

The current configs are for the blender, LLFF, and shapenet datasets, which can be found in configs.

After setting the parameters of the model, to run it,

python run.py --configs/config_DATATYPE

Practical Concerns

The models were tested on 32gb GPUs, and higher resolution images require very large amounts of memory. The shapenet experiments should run on 16gb GPUs.

Acknowledgements

The code is built upon the original NeRF implementation. Thanks to LucidRains for the torch implementation of slot attention on which the current version is based.

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{grf2020,
  title={GRF: Learning a General Radiance Field for 3D Scene Representation and Rendering},
  author={Trevithick, Alex and Yang, Bo},
  booktitle={arXiv:2010.04595},
  year={2020}
}
Owner
Alex Trevithick
ML + CV👍
Alex Trevithick
Language models are open knowledge graphs ( non official implementation )

language-models-are-knowledge-graphs-pytorch Language models are open knowledge graphs ( work in progress ) A non official reimplementation of Languag

theblackcat102 132 Dec 18, 2022
A fast implementation of bss_eval metrics for blind source separation

fast_bss_eval Do you have a zillion BSS audio files to process and it is taking days ? Is your simulation never ending ? Fear no more! fast_bss_eval i

Robin Scheibler 99 Dec 13, 2022
Video Autoencoder: self-supervised disentanglement of 3D structure and motion

Video Autoencoder: self-supervised disentanglement of 3D structure and motion This repository contains the code (in PyTorch) for the model introduced

157 Dec 22, 2022
I explore rock vs. mine prediction using a SONAR dataset

I explore rock vs. mine prediction using a SONAR dataset. Using a Logistic Regression Model for my prediction algorithm, I intend on predicting what an object is based on supervised learning.

Jeff Shen 1 Jan 11, 2022
The Video-based Accident Detection System built in Python

Accident-detection-system About the Project This Repository contains the Video-based Accident Detection System built in Python. Contributors Yukta Gop

SURYAVANSHI SNEHAL BALKRISHNA 50 Dec 07, 2022
Pytorch implementation of few-shot semantic image synthesis

Few-shot Semantic Image Synthesis Using StyleGAN Prior Our method can synthesize photorealistic images from dense or sparse semantic annotations using

40 Sep 26, 2022
Code and results accompanying our paper titled Mixture Proportion Estimation and PU Learning: A Modern Approach at Neurips 2021 (Spotlight)

Mixture Proportion Estimation and PU Learning: A Modern Approach This repository is the official implementation of Mixture Proportion Estimation and P

Approximately Correct Machine Intelligence (ACMI) Lab 23 Dec 28, 2022
Stochastic Scene-Aware Motion Prediction

Stochastic Scene-Aware Motion Prediction [Project Page] [Paper] Description This repository contains the training code for MotionNet and GoalNet of SA

Mohamed Hassan 31 Dec 09, 2022
Code for "On Memorization in Probabilistic Deep Generative Models"

On Memorization in Probabilistic Deep Generative Models This repository contains the code necessary to reproduce the experiments in On Memorization in

The Alan Turing Institute 3 Jun 09, 2022
Official implementation of the paper "AAVAE: Augmentation-AugmentedVariational Autoencoders"

AAVAE Official implementation of the paper "AAVAE: Augmentation-AugmentedVariational Autoencoders" Abstract Recent methods for self-supervised learnin

Grid AI Labs 48 Dec 12, 2022
PyTorch implementation of "MLP-Mixer: An all-MLP Architecture for Vision" Tolstikhin et al. (2021)

mlp-mixer-pytorch PyTorch implementation of "MLP-Mixer: An all-MLP Architecture for Vision" Tolstikhin et al. (2021) Usage import torch from mlp_mixer

isaac 27 Jul 09, 2022
Learning Tracking Representations via Dual-Branch Fully Transformer Networks

Learning Tracking Representations via Dual-Branch Fully Transformer Networks DualTFR ⭐ We achieves the runner-ups for both VOT2021ST (short-term) and

phiphi 19 May 04, 2022
The missing CMake project initializer

cmake-init - The missing CMake project initializer Opinionated CMake project initializer to generate CMake projects that are FetchContent ready, separ

1k Jan 01, 2023
Namish Khanna 40 Oct 11, 2022
Efficient Householder transformation in PyTorch

Efficient Householder Transformation in PyTorch This repository implements the Householder transformation algorithm for calculating orthogonal matrice

Anton Obukhov 49 Nov 20, 2022
Course materials for Fall 2021 "CIS6930 Topics in Computing for Data Science" at New College of Florida

Fall 2021 CIS6930 Topics in Computing for Data Science This repository hosts course materials used for a 13-week course "CIS6930 Topics in Computing f

Yoshi Suhara 101 Nov 30, 2022
Code for our CVPR 2021 paper "MetaCam+DSCE"

Joint Noise-Tolerant Learning and Meta Camera Shift Adaptation for Unsupervised Person Re-Identification (CVPR'21) Introduction Code for our CVPR 2021

FlyingRoastDuck 59 Oct 31, 2022
[CVPR 2022] Structured Sparse R-CNN for Direct Scene Graph Generation

Structured Sparse R-CNN for Direct Scene Graph Generation Our paper Structured Sparse R-CNN for Direct Scene Graph Generation has been accepted by CVP

Multimedia Computing Group, Nanjing University 44 Dec 23, 2022
curl-impersonate: A special compilation of curl that makes it impersonate Chrome & Firefox

curl-impersonate A special compilation of curl that makes it impersonate real browsers. It can impersonate the four major browsers: Chrome, Edge, Safa

lwthiker 1.9k Jan 03, 2023
CAUSE: Causality from AttribUtions on Sequence of Events

CAUSE: Causality from AttribUtions on Sequence of Events

Wei Zhang 21 Dec 01, 2022