HiCMAE: Hierarchical Contrastive Masked Autoencoder for Self-Supervised Audio-Visual Emotion Recognition
[arXiv] | [Information Fusion]
Licai Sun, Zheng Lian, Bin Liu, and Jianhua Tao
University of Chinese Academy of Sciences & Institute of Automation, Chinese Academy of Sciences & Tsinghua University
📰 News
[2024.10.21] We upload the fine-tuned models on CREMA-D and MAFW.
[2024.04.11] We upload the pre-training code.
[2024.03.20] Our paper is accepted by Information Fusion.
[2024.01.11] We upload the initial code and pre-trained model.
✨ Overview
Abstract: Audio-Visual Emotion Recognition (AVER) has garnered increasing attention in recent years for its critical role in creating emotion-ware intelligent machines. Previous efforts in this area are dominated by the supervised learning paradigm. Despite significant progress, supervised learning is meeting its bottleneck due to the longstanding data scarcity issue in AVER. Motivated by recent advances in self-supervised learning, we propose Hierarchical Contrastive Masked Autoencoder (HiCMAE), a novel self-supervised framework that leverages large-scale self-supervised pre-training on vast unlabeled audio-visual data to promote the advancement of AVER. Following prior arts in self-supervised audio-visual representation learning, HiCMAE adopts two primary forms of self-supervision for pre-training, namely masked data modeling and contrastive learning. Unlike them which focus exclusively on top-layer representations while neglecting explicit guidance of intermediate layers, HiCMAE develops a three-pronged strategy to foster hierarchical audio-visual feature learning and improve the overall quality of learned representations. Firstly, it incorporates hierarchical skip connections between the encoder and decoder to encourage intermediate layers to learn more meaningful representations and bolster masked audio-visual reconstruction. Secondly, hierarchical cross-modal contrastive learning is also exerted on intermediate representations to narrow the audio-visual modality gap progressively and facilitate subsequent cross-modal fusion. Finally, during downstream fine-tuning, HiCMAE employs hierarchical feature fusion to comprehensively integrate multi-level features from different layers. To verify the effectiveness of HiCMAE, we conduct extensive experiments on 9 datasets covering both categorical and dimensional AVER tasks. Experimental results show that our method significantly outperforms state-of-the-art supervised and self-supervised audio-visual methods, which indicates that HiCMAE is a powerful audio-visual emotion representation learner.
🚀 Main Results
Comparison with state-of-the-art audio-visual methods on 9 datasets.
Please check our arXiv paper to see detailed results on each dataset.
👀 Visualization
✨ Reconstruction
Please check our arXiv paper to see details better.
✨ t-SNE on CREMA-D
🔨 Installation
Main prerequisites:
Python 3.8PyTorch 1.10.1 (cuda 11.3), torchvision==0.11.2, torchaudio==0.10.1timm==0.4.12einops==0.6.1decord==0.6.0openmim==0.3.6, mmcv==1.7.1scikit-learn=1.2.1, scipy=1.10.0, pandas==1.5.3, numpy=1.23.5opencv-python=4.7.0.72tensorboardX=2.6.1soundfile==0.12.1
If some are missing, please refer to requirements.txt for more details.
➡️ Data Preparation
-
If the original dataset does not provide extracted faces, we use OpenFace to extract them from videos. Please see extract_face.py in preprocess for details.
-
If the original dataset does not provide extracted audios (sampling rate: 16k), we use FFmpeg to extract them from videos. Please see extract_audio.py in preprocess for details.
-
Follow the files (e.g., cremad_av.py) in preprocess to prepare your annotation files.
Specifically, you need to enerate annotations for dataloader ("
" in annotations). The annotation usually includes train.csvandtest.csv. The format of*.csvfile is typically like this:dataset_root/video_1 dataset_root/audio_1 label_1 dataset_root/video_2 dataset_root/audio_2 label_2 dataset_root/video_3 dataset_root/audio_3 label_3 ... dataset_root/video_N dataset_root/audio_N label_NAn example of train.csv for CREMA-D (6-class) fold1 (fd1) is shown as follows:
/data/ycs/AC/Dataset/CREMA-D/face_aligned/1019_DFA_ANG_XX /data/ycs/AC/Dataset/CREMA-D/AudioWAV/1019_DFA_ANG_XX.wav 0 /data/ycs/AC/Dataset/CREMA-D/face_aligned/1019_DFA_DIS_XX /data/ycs/AC/Dataset/CREMA-D/AudioWAV/1019_DFA_DIS_XX.wav 1 /data/ycs/AC/Dataset/CREMA-D/face_aligned/1019_DFA_FEA_XX /data/ycs/AC/Dataset/CREMA-D/AudioWAV/1019_DFA_FEA_XX.wav 2
📍Pre-train HiCMAE
-
VoxCeleb2
sh scripts/voxceleb2/audio_visual/hicmae_pretrain_base/pretrain_base.shYou can download our pre-trained model on VoxCeleb2 from this link and put it into this folder.
⤴️ Fine-tuning with pre-trained models
-
CREMA-D
sh scripts/voxceleb2/audio_visual/hicmae_pretrain_base/cream-d/finetune_170.shThe fine-tuned checkpoints and logs across five folds on CREMA-D are provided as follows: Fold UAR WR Fine-tuned Model 1 86.66 86.67 log / checkpoint 2 83.27 83.20 log / checkpoint 3 87.23 87.19 log / checkpoint 4 83.70 83.79 log / checkpoint 5 83.88 83.79 log / checkpoint Total 84.94 84.91 - -
MAFW
sh scripts/voxceleb2/audio_visual/hicmae_pretrain_base/mafw/finetune_170.shThe fine-tuned checkpoints and logs across five folds on MAFW are provided as follows: Fold UAR WR Fine-tuned Model 1 36.02 47.60 log / checkpoint 2 41.57 55.10 log / checkpoint 3 46.46 60.10 log / checkpoint 4 47.50 63.09 log / checkpoint 5 41.88 55.22 log / checkpoint Total 42.69 56.21 -
☎️ Contact
If you have any questions, please feel free to reach me out at sunlicai2019@ia.ac.cn.
👍 Acknowledgements
This project is built upon VideoMAE and AudioMAE. Thanks for their great codebase.
✏️ Citation
If you think this project is helpful, please feel free to leave a star⭐️ and cite our paper:
@article{sun2024hicmae,
title={HiCMAE: Hierarchical Contrastive Masked Autoencoder for Self-Supervised Audio-Visual Emotion Recognition},
author={Sun, Licai and Lian, Zheng and Liu, Bin and Tao, Jianhua},
journal={arXiv preprint arXiv:2401.05698},
year={2024}
}