An Open-source Benchmark of Deep Learning Models for Audio-visual Apparent and Self-reported Personality Recognition

Introduction

This is the official code repo of An Open-source Benchmark of Deep Learning Models for Audio-visual Apparent and Self-reported Personality Recognition (https://arxiv.org/abs/2210.09138).

In this project, seven visual models, six audio models and five audio-visual models have been reproduced and evaluated. Besides, seven widely-used visual deep learning models, which have not been applied to video-based personality computing before, have also been employed for benchmark. Detailed description can be found in our paper.

All benchmarked models are evaluated on: the ChaLearn First Impression dataset and the ChaLearn UDIVA self-reported personality dataset

This project is currently under active development. Documentation, examples, and tutorial will be progressively detailed

Installation

Setup project: you can use either Conda or Virtualenv/pipenv to create a virtual environment to run this program.

# create and activate a virtual environment
virtualenv -p python38 venv
source venv/bin/activate

Installing from PyPI

pip install deep_personality

Installing from Github

# clone current repo
git clone DeepPersonality
cd DeepPersonality

# install required packages and dependencies
pip install -r requirements.txt

Datasets

The datasets we used for benchmark are Chalearn First Impression and UDIVA.

The former contains 10, 000 video clips that come from 2, 764 YouTube users for apparent personality recognition(impression), where each video lasts for about 15 seconds with 30 fps.
The latter, for self-reported personality, records 188 dyadic interaction video clips between 147 voluntary participants, with total 90.5h of recordings. Every clip contains two audiovisual files, where each records a single participant’s behaviours.
Each video in both datasets is labelled with the Big-Five personality traits.

To meet various requirements from different models and experiments, we extract raw audio file and all frames from a video and then extract face images from each full frame, termed as face frames.

For quick start and demonstration, we provide a tiny Chalearn 2016 dataset containing 100 videos within which 60 for training, 20 for validation and 20 for test. Please find the process methods in dataset preparation.

For your convenience, we provide the processed face image frames dataset for Chalearn 2016 since that dataset is publicly available, which indicates we can make our processed data open to the community.

Experiments

Reproducing reported experiments

We employ a build-from-config manner to conduct an experiment. After setting up the environments and preparing the data needed, we can have a quick start by the following command line:

# cd DeepPersonality # top directory 
script/run_exp.py --config path/to/exp_config.yaml

For quick start with tiny ChaLearn 2016 dataset, if you prepare the data by the instructions in above section, the following command will launch an experiment for bimodal-resnet18 model.

# cd DeepPersonality # top directory
script/run_exp.py --config config/demo/bimodal_resnet18.yaml

Detailed arguments description are presented in command line interface file.

For quick start demonstration, please find the Colab Notebook: QuickStart

For experiments start from raw video processing, please find this Colab Notebook: StartFromDataProcessing

Developing new personality computing models

We use config-pipe line files and registration mechanism to organize our experiments. If user want to add their own models or algorithms into this program please reference the Colab Notebook TrainYourModel

Models

On ChaLearn 2016 dataset

Model	Modal	ChaLearn2016 cfgs	ChaLearn2016 weights
DAN	visual	cfg	weight
CAM-DAN⁺	visual	cfg	weight
ResNet	visual	cfg	weight
HRNet	visual	cfg-frame/cfg-face	weight-frame/weight-face
SENet	visual	cfg-frame/cfg-face	weight
3D-ResNet	visual	cfg-frame/cfg-face	weight
Slow-Fast	visual	cfg-frame/cfg-face	weight
TPN	visual	cfg-frame/cfg-face	weight
Swin-Transformer	visual	cfg-frame/cfg-face	weight
VAT	visual	cfg-frame/cfg-face	weight
Interpret Audio CNN	audio	cfg	weight
Bi-modal CNN-LSTM	audiovisual	cfg	weight
Bi-modal ResNet	audiovisual	cfg	weight
PersEmoN	audiovisual	cfg	weight
CRNet	audiovisual	cfg	weight
Amb-Fac	audiovisual	cfg-frame, cfg-face	weight

On ChaLearn 2021 dataset

Model	Modal	Talk	Lego	Ghost	Animal
DAN	visual	cfg weight	cfg weight	cfg weight	cfg weight
CAM-DAN⁺	visual	cfg weight	cfg weight	cfg weight	cfg weight
ResNet	visual	cfg weight	cfg weight	cfg weight	cfg weight
HRNet	visual	cfg weight	cfg weight	cfg weight	cfg weight
SENet	visual	cfg weight	cfg weight	cfg weight	cfg weight
3D-ResNet	visual	cfg weight	cfg weight	cfg weight	cfg weight
Slow-Fast	visual	cfg weight	cfg weight	cfg weight	cfg weight
TPN	visual	cfg weight	cfg weight	cfg weight	cfg weight
Swin-Transformer	visual	cfg weight	cfg weight	cfg weight	cfg weight
VAT	visual	cfg weight	cfg weight	cfg weight	cfg weight
Interpret Audio CNN	audio	cfg weight	cfg weight	cfg weight	cfg weight
Bi-modal CNN-LSTM	audiovisual	cfg weight	cfg weight	cfg weight	cfg weight
Bi-modal ResNet	audiovisual	cfg weight	cfg weight	cfg weight	cfg weight
PersEmoN	audiovisual	cfg weight	cfg weight	cfg weight	cfg weight
CRNet	audiovisual	cfg weight	cfg weight	cfg weight	cfg weight
Amb-Fac	audiovisual	cfg weight	cfg weight	cfg weight	cfg weight

Papers

From which the models are reproduced

Deep bimodal regression of apparent personality traits from short video sequences
Bi-modal first impressions recognition using temporally ordered deep audio and stochastic visual features
Deep impression: Audiovisual deep residual networks for multimodal apparent personality trait recognition
Cr-net: A deep classification-regression network for multimodal apparent personality analysis
Interpreting cnn models for apparent personality trait regression
On the use of interpretable cnn for personality trait recognition from audio
Persemon: a deep network for joint analysis of apparent personality, emotion and their relationship
A multi-modal personality prediction system
Squeeze-and-excitation networks
Deep high-resolution representation learning for visual recognition
Swin transformer: Hierarchical vision transformer using shifted windows
Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet
Slowfast networks for video recognition
Temporal pyramid network for action recognition
Video action transformer network

Citation

If you use our code for a publication, please kindly cite it as:

@article{liao2022open,
  title={An Open-source Benchmark of Deep Learning Models for Audio-visual Apparent and Self-reported Personality Recognition},
  author={Liao, Rongfan and Song, Siyang and Gunes, Hatice},
  journal={arXiv preprint arXiv:2210.09138},
  year={2022}
}

History

2022/10/17 - Paper submission and make project publicly available.

To Be Updated

Test Pip install
Description of adding new models
Model zoo
Notebook tutorials

liaorongfan / DeepPersonality