AudioEar: Single-View Ear Reconstruction for Personalized Spatial Audio

This is the official code of AAAI'23 paper AudioEar: Single-View Ear Reconstruction for Personalized Spatial Audio written in PyTorch.

Note: This codebase is a naive version and is under refactoring. If you have any questions, feel free to contact the author.

Paper

Introduction

Spatial audio, which focuses on immersive 3D sound rendering, is widely applied in the acoustic industry. One of t he key problems of current spatial audio rendering methods is the lack of personalization based on different anatomies of individuals, which is essential to produce accurate sound source positions. In this work, we address this problem from an interdisciplinary perspective. The rendering of spatial audio is strongly correlated with the 3D shape of human bodies, particularly ears.

To this end, we propose to achieve personalized spatial audio by reconstructing 3D human ears with single-view images. First, to benchmark the ear reconstruction task, we introduce AudioEar3D, a high-quality 3D ear dataset consisting of 112 point cloud ear scans with RGB images. To self-supervisedly train a reconstruction model, we further collect a 2D ear dataset composed of 2,000 images, each one with manual annotation of occlusion and 55 landmarks, named AudioEar2D. To our knowledge, both datasets have the largest scale and best quality of their kinds for public use. Further, we propose AudioEarM, a reconstruction method guided by a depth estimation network that is trained on synthetic data, with two loss functions tailored for ear data. Lastly, to fill the gap between the vision and acoustics community, we develop a pipeline to integrate the reconstructed ear mesh with an off-the-shelf 3D human body and simulate a personalized Head-Related Transfer Function (HRTF), which is the core of spatial audio rendering.

Published Datasets

Our collected datasets, AudioEar3D and AudioEar2D, could be downloaded in Zenodo. If you find any downloading problem, you could also use this mirror link from Google Drive.

AudioEar3D

3D Ear Dataset	Scale	with Image	Quality	Accessibility
UND-J2	1,800	✔	*	✔
York3DEar	500	✘	*	✔
SYMARE-1	20	✘	***	✔
SYMARE-2	102	✘	***	✘
Ploumpis et al.	234	✘	***	✘
AudioEar3D	112	✔	****	✔

AudioEar2D

2D Ear Dataset	Scale	Source	Landmark Annotations	Usage
UND-E	464	Limited	✘	Biometrics
AMI	700	Limited	✘	Biometrics
IIT Delhi Ear	754	Limited	✘	Biometrics
WPUTEDB	3348	Limited	✘	Biometrics
UBEAR	4,410	Limited	✘	Biometrics
IBug-B	2,058	In-the-wild	✘	Biometrics
AWE	9,500	In-the-wild	✘	Biometrics
EarVN	28,412	In-the-wild	✘	Biometrics
IBug-A	605	In-the-wild	✔	Reconstruction
AudioEar2D	2,000	In-the-wild	✔	Reconstruction

Data Preparation

1. Prepare AudioEar2D Dataset

   • Download AudioEar2D dataset, then modify the cfg.model.ear_dataset_path in the config.py to the directory of the dataset.
   • Copy the train/test split file split.json file in ./data/AudioEar2D to the dataset folder.

2. Prepare Texture Model

   • Follow the instructions for the Albedo model to get 'FLAME_albedo_from_BFM.npz', then modify the cfg.model.tex_path in the config.py to the path of the npz file.

3. Prepare AudioEar3D Dataset

   • Download AudioEar3D dataset, then modify the cfg.s2m.s2m_data_path in the config.py to the directory of the dataset.

4. Prepare Synthetic Dataset

   • Download Synthetic dataset from here, then modify the cfg.s2m.sythetic_dataset_path in the config.py to the directory of the dataset.

Pre-trained Weights

The pre-train weights of our ResNet encoder and Monocular Depth Estimation (MDE) model could downloaded through this link.

Requirements

• Python 3.7

• PyTorch>=1.6

• PyTorch3D>=0.7.0

• CUDA Toolkit>10.2

• Trimesh>=3.9

• numpy>=1.18.5

• scipy>=1.4.1

• chumpy>=0.69

• scikit-image>=0.15

• opencv-python>=4.1.1

• scikit-image>=0.15

• PyYAML>=5.1.1

  You can install them either manually or through the command:

  pip install -r requirements.txt

  Pytorch3D might need manual installation. Follow the official instruction to install it.

Usage

1. Training and validating MDE model on Synthetic dataset

   python train_depth.py

2. Training and validating on AudioEar2D dataset

   Modify the cfg.model.depth_model_path to the model obtained from train_depth.py. Then run:

   python train_recon.py

3. Evaluation on AudioEar3D dataset

   Change the cfg.s2m.recon_model_path in the config.py to the model checkpoint file obtained from train_recon.py . Then run:

   python s2mdemo.py
   

Citation

If you find this project useful in your research, please cite the paper as:

Xiaoyang Huang, Yanjun Wang, Yang Liu, Bingbing Ni, Wenjun Zhang, Jinxian Liu, Teng Li. "AudioEar: Single-View Ear Reconstruction for Personalized Spatial Audio". arXiv preprint arXiv:2301.12613, 2023.

or using bibtex:

@article{huang2023audioear,
  title={AudioEar: Single-View Ear Reconstruction for Personalized Spatial Audio},
  author={Huang, Xiaoyang and Wang, Yanjun and Liu, Yang and Ni, Bingbing and Zhang Wenjun and Liu Jinxian and Li, Teng},
  journal={arXiv preprint arXiv:2301.12613},
  year={2023}
}