Neural Articulated Radiance Field
NARF
Neural Articulated Radiance Field
Atsuhiro Noguchi, Xiao Sun, Stephen Lin, Tatsuya Harada
ICCV 2021
Abstract
We present Neural Articulated Radiance Field (NARF), a novel deformable 3D representation for articulated objects learned from images. While recent advances in 3D implicit representation have made it possible to learn models of complex objects, learning pose-controllable representations of articulated objects remains a challenge, as current methods require 3D shape supervision and are unable to render appearance. In formulating an implicit representation of 3D articulated objects, our method considers only the rigid transformation of the most relevant object part in solving for the radiance field at each 3D location. In this way, the proposed method represents pose-dependent changes without significantly increasing the computational complexity. NARF is fully differentiable and can be trained from images with pose annotations. Moreover, through the use of an autoencoder, it can learn appearance variations over multiple instances of an object class. Experiments show that the proposed method is efficient and can generalize well to novel poses.
Method
We extend Neural Radiance Fields (NeRF) to articulated objects. NARF is a NeRF conditioned on skeletal parameters and skeletal posture, and is an MLP that outputs the density and color of a point with 3D position and 2D viewing direction as input. Since articulated objects can be regarded as multiple rigid bodies connected by joints, the following two assumptions can be made
- The density of each part does not change in the coordinate system fixed to the part.
- A point on the surface of the object belongs to only one of the parts.
Therefore, we transform the input 3D coordinates into local coordinates of each part and use them as input for the model. From the second hypothesis, we use selector MLP to select only one necessary coordinate and mask the others.
An overview of the model is shown in the figure.
The model is trained with the L2 loss between the generated image and the ground truth image.
Results
The proposed NARF is capable of rendering images with explicit control of the viewpoint, bone pose, and bone parameters. These representations are disentangled and can be controlled independently.
Viewpoint change (seen in training)
Pose change (unseen in training)
Bone length change (unseen in training)
NARF generalizes well to unseen viewpoints during training.
Furthermore, NARF can render segmentation for each part by visualizing the output values of the selector.
NARF can learn appearance variations by combining it with an autoencoder. The video below visualizes the disentangled representations and segmentation masks learned by NARF autoencoder.
Code
Envirionment
python 3.7.*
pytorch >= 1.7.1
torchvision >= 0.8.2
pip install tensorboardx pyyaml opencv-python pandas ninja easydict tqdm scipy scikit-image
Dataset preparation
THUman
Please refer to https://github.com/nogu-atsu/NARF/tree/master/data/THUman
Your own dataset
Coming soon.
Training
-
Write config file like
NARF/configs/THUman/results_wxl_20181008_wlz_3_M/NARF_D.yml
. Do not changedefault.yml
out_root
: root directory to save modelsout
: experiment namedata_root
: directory thedataset
is in
-
Run training specifying a config file
CUDA_VISIBLE_DEVICES=0 python train.py --config NARF/configs/[your_config.yml] --num_workers 1
-
Distributed data parallel
python train_ddp.py --config NARF/configs/[your_config.yml] --gpus 4 --num_workers 1
Validation
-
Single gpu
python train.py --config NARF/configs/[your_config.yml] --num_workers 1 --validation --resume_latest
-
Multiple gpus
python train_ddp.py --config NARF/configs/[your_config.yml] --gpus 4 --num_workers 1 --validation --resume_latest
-
The results are saved to
val_metrics.json
in the same directory as the snapshots.
Computational cost
python computational_cost.py --config NARF/configs/[your_config.yml]
Visualize results
-
Generate interpolation videos
cd visualize python NARF_interpolation.py --config ../NARF/configs/[your_config.yml]
The results are saved to the same directory as the snapshots. With the default settings, it takes 30 minutes on a V100 gpu to generate a 30-frame video
Acknowledgement
https://github.com/rosinality/stylegan2-pytorch
https://github.com/ZhengZerong/DeepHuman
https://smpl.is.tue.mpg.de/
BibTex
@inproceedings{2021narf,
author = {Noguchi, Atsuhiro and Sun, Xiao and Lin, Stephen and Harada, Tatsuya},
title = {Neural Articulated Radiance Field},
booktitle = {International Conference on Computer Vision},
year = {2021},
}