Bridging the Gap: Protocol Towards Fair and Consistent Affect Analysis

The project is an official implementation of our paper:

Bridging the Gap: Protocol Towards Fair and Consistent Affect Analysis (IEEE FG 2024)

Guanyu Hu, Eleni Papadopoulou, Dimitrios Kollias, Paraskevi Tzouveli, Jie Wei and Xinyu Yang

The increasing integration of machine learning algorithms in daily life underscores the critical need for fairness and equity in their deployment. As these technologies play a pivotal role in decision-making, addressing biases across diverse subpopulation groups, including age, gender, and race, becomes paramount. Automatic affect analysis, at the intersection of physiology, psychology, and machine learning, has seen significant development. However, existing databases and methodologies lack uniformity, leading to biased evaluations. This work addresses these issues by analyzing six affective databases, annotating demographic attributes, and proposing a common protocol for database partitioning. Emphasis is placed on fairness in evaluations. Extensive experiments with baseline and state-of-the-art methods demonstrate the impact of these changes, revealing the inadequacy of prior assessments. The findings underscore the importance of considering demographic attributes in affect analysis research and provide a foundation for more equitable methodologies.

1 Access to Developments of our Work

If you are an academic (i.e., a person with a permanent position at a university, e.g. a professor, but not a Post-Doc or a PhD/PG/UG student), to request the developments of our work (i.e., annotations, partitions, and pre-trained models-checkpoints), please:
i) send an email to d.kollias@qmul.ac.uk with subject: Fair and Consistent Affect Analysis request by academic;
ii) use your official academic email (as data cannot be released to personal emails);
iii) include in the email the AGREEMENT FORM, the reason why you require access to these developments, and your official academic website.

If you are from industry and you want to acquire the developments of our work (i.e., annotations, partitions, and pre-trained models-checkpoints), please send an email from your official industrial email to d.kollias@qmul.ac.uk with subject: Fair and Consistent Affect Analysis request from industry and explain the reason why the database access is needed; also specify if it is for research or commercial purposes.

2 Datasets

2.1 AffectNet 📷 - EXPR & VA

You can download the dataset from http://mohammadmahoor.com/affectnet/

EXPR-7/8

• New Partition and Annotations (5 columns):
  • The split between AffectNet 7 and AffectNet 8 is exactly the same. By simply deleting the images and corresponding labels for 'contempt', AffectNet 7 can be derived from AffectNet 8
  • 8 class CSV Path: AffectNet/annotations/EXPR
    • Expression Labels: ['Anger', 'Contempt', 'Disgust', 'Fear', 'Happiness', 'Neutral', 'Sadness', 'Surprise']
  • Header: [name, expression, age, gender, race]
  • Demographic Labels:
    • Gender: [Female, Male]
    • Age: [00-02, 03-09, 10-19, 20-29, 30-39, 40-49, 50-59, 60-69, 70+]
    • Race: [White, Black, Asian, Indian]

VA

• New Partition and Annotations:
  • CSV Path: AffectNet/annotations/VA
  • Header: [name, valence, arousal, age, gender, race]
  • Demographic Labels:
    • Gender: [Female, Male]
    • Age: [00-02, 03-09, 10-19, 20-29, 30-39, 40-49, 50-59, 60-69, 70+]
    • Race: [White, Black, Asian, Indian]

2.2 DISFA 🎥 - AU

You can download the dataset from http://mohammadmahoor.com/disfa/

• New Partition and Annotations:
  • 12 AUs
  • CSV Path: DISFA/annotations
  • Header: [name, AU1, AU2, AU4, AU5, AU6, AU9, AU12, AU15, AU17, AU20, AU25, AU26, age, gender, race]
    • AU Labels: [1: present, 0: absent]
  • Demographic Labels:
    • Gender: [Female, Male]
    • Age: [10-19, 20-29, 30-39, 40-49]]
    • Race: [White, Black, Asian, Indian]

2.3 EmotioNet 📷 - AU

You can download the dataset from https://cbcsl.ece.ohio-state.edu/dbform_emotionet.html

• New Partition and Annotations:
  • 11 AUs
  • CSV Path: EmotioNet/annotation
  • Header: [name, AU1, AU2, AU4, AU5, AU6, AU9, AU12, AU17, AU20, AU25, AU26, age, gender, race]
  • AU: [1: present, 0: absent, -1: unknown]
  • Demographic Labels:
    • Gender: [Female, Male]
    • Age: [00-02, 03-09, 10-19, 20-29, 30-39, 40-49, 50-59, 60-69, 70+]
    • Race: [White, Black, Asian, Indian]

2.4 GFT 🎥 - AU

You can download the dataset from https://osf.io/7wcyz/

• New Partition and Annotations:
  • 14 AUs
  • Path: RAF_AU/annotation
  • Header: [name, AU1, AU2, AU4, AU5, AU6, AU7, AU9, AU10, AU11, AU12, AU15, AU17, AU23, AU24, age, gender, race]
  • Demographic Labels:
    • Gender: [Female, Male]
    • Age: [21, 22, 23, 24, 25, 26, 27, 28]
    • Race: [White, Black, Asian, Other]

2.5 RAF-DB 📷 - EXPR

You can download the dataset from http://www.whdeng.cn/raf/model1.html

• New Partition and Annotations:
  • 7 Expressions
  • CSV Path: RAF_DB/annotation
  • Header: [name, gender, race, age, expression]
    • Expression Labels: ['Anger', 'Disgust', 'Fear', 'Happiness', 'Neutral', 'Sadness', 'Surprise']
  • Demographic Labels:
    • Gender: [Female, Male, Unsure]
    • Age: [00-03, 04-19, 20-39, 40-69, 70+]
    • Race: [White, Black, Asian, Indian]

2.6 RAF-AU 📷 - AU

You can download the dataset from http://whdeng.cn/RAF/model3.html

• New Partition and Annotations:
  • 13 AUs
  • CSV Path: RAF_AU/annotation
  • Header: [name, AU1, AU2, AU4, AU5, AU6, AU9, AU10, AU12, AU16, AU17, AU25, AU26, AU27, age, gender, race]
  • Demographic Labels:
    • Gender: [Female, Male]
    • Age: [00-02, 03-09, 10-19, 20-29, 30-39, 40-49, 50-59, 60-69, 70+]
    • Race: [White, Black, Asian, Indian]

3 Models

Support the following models:

Model Type	Models
ResNet	resnet18, resnet34, resnet50, resnet101, resnet152
ResNeXt	resnext50_32x4d, resnext101_32x8d, resnext101_64x4d
Swin	swin_t, swin_s, swin_b, swin_v2_t, swin_v2_s, swin_v2_b
VGG	vgg11, vgg16, vgg19
ViT	vit_b_16, vit_b_32, vit_l_16, vit_l_32, vit_h_14
iResNet	iresnet101
EfficientNet	efficientnet_b0, efficientnet_b1, efficientnet_b2, efficientnet_b6, efficientnet_b7, efficientnet_v2_s, efficientnet_v2_m, efficientnet_v2_l
DenseNet	densenet121, densenet161, densenet201
ConvNeXt	convnext_tiny, convnext_small, convnext_base, convnext_large
POSTER++	POSTER++
DAN	DAN
MT-EffNet	MT-EffNet
MA-Net	MA-Net
EAC	EAC
DACL	DACL
EmoGCN	EmoGCN
CTC	abaw5_ctc
SITU	abaw5_situ
FUXI	FUXI
ME-GraphAU	ME-GraphAU
AUNets	AUNets

4 Requirements

Our experiments are conducted in the following environments:

• Python 3.11
• PyTorch 2.3
• CUDA 11.8
• Torchvision 0.18

You can set up your environment by following the instructions in ENVIRONMENT.md.

5 Code Structure

Fair-Consistent-Affect-Analysis   # Project code root
├── config                        # Configuration and parameters
├── fair                          # Fairness metrics
├── model_utils                   # Model utilities
├── ENVIRONMENT.md                # Requirements
├── run.py                        # Main
├── output                        # Created automatically
└── README.md                     # Project readme

6 Data Preparation

Your data needs to be structured in the following format:

Dataset_root                   # dataset root
├── AffectNet                  # AffectNet
│   ├── images
│   │   ├── train_set
│   │   │   ├── 0.jpg
│   │   │   └── ...
│   │   └── val_set
│   │       ├── 0.jpg
│   │       └── ...
│   └── annotations
│       ├── EXPR
│       │   ├── train.csv
│       │   ├── valid.csv
│       │   └── test.csv
│       └── VA
│           ├── train.csv
│           ├── valid.csv
│           └── test.csv
├── DISFA                      # DISFA
│   ├── images
│   │   ├── LeftVideoSN001
│   │   │   ├── 00001.jpg
│   │   │   └── ...
│   │   └── RightVideoSN001
│   │       ├── 00001.jpg
│   │       └── ...
│   └── annotations
│       ├── train.csv
│       ├── valid.csv
│       └── test.csv
├── EmotioNet                  # EmotioNet
│   ├── images
│   │   ├── test
│   │   │   ├── 000001.jpg
│   │   │   └── ...
│   │   └── validation
│   │       ├── 000001.jpg
│   │       └── ...
│   └── annotations
│       ├── train.csv
│       ├── valid.csv
│       └── test.csv
├── GFT                        # GFT
│   ├── images
│   │   ├── 001A
│   │   │   ├── 1.jpg
│   │   │   └── ...
│   │   └── 001B
│   │       ├── 1.jpg
│   │       └── ...
│   └── annotations
│       ├── train.csv
│       ├── valid.csv
│       └── test.csv
├── RAF-AU                     # RAF-AU
│   ├── images
│   │   ├── 0001_aligned.jpg
│   │   └── ...
│   └── annotations
│       ├── train.csv
│       ├── valid.csv
│       └── test.csv
└── RAF-DB                     # RAF-DB
    ├── images
    │   ├── test_0001.jpg
    │   ├── ...
    │   ├── train_00001.jpg
    │   └── test.csv
    └── annotations
        ├── train.csv
        ├── valid.csv
        └── test.csv

7 Config

To train or test, you need to modify the root path in ./config/config.yaml to your own path.

Root Path

dataset_root: &dataset_root /path/to/Dataset_root
restructured_csv_root: &restructured_csv_root /path/to/Dataset_root
output_root: /path/to/Fair-Consistent-Affect-Analysis
code_root: /path/to/Fair-Consistent-Affect-Analysis

Dataset Path

If you structured your dataset repository following the Data Preparation guidelines, you do not need to change any content below. Otherwise, you need to change the paths to match your own directory structure.

• [Dataset Name]_csv_root: Dataset CSV folder, where CSV files must follow the annotations folder structure as described in the Data Preparation section.
• [Dataset Name]_num_class: For expression datasets, it should be the number of expressions. For AU datasets, it should be the number of AUs. For VA datasets, it should be set to 2.

The following is an example for the dataset AffectNet-7:

AffectNet-7_dataset_root: !join_path [ *dataset_root, AffectNet/images ]  # Dataset folder
AffectNet-7_csv_root: !join_path [ *restructured_csv_root, AffectNet/annotations/EXPR/expr7 ]  # Dataset CSV folder
AffectNet-7_num_class: 7  # Number of classes

8 Training

Training Parameters

• dataset: Controls the training dataset. The following datasets are available:
  • EXPR: AffectNet-7, AffectNet-8, RAF-DB
  • AU: EmotioNet, GFT, RAF-AU, DISFA
  • VA: AffectNet-VA
• model: Controls the training model. The models available can be found in the Models section.
• fair: Run fair validation.

Training Examples

• To run training on the dataset AffectNet-7 using the model resnet18:

python run.py --dataset AffectNet-7 --model resnet18 -bs 128 --fair

• To run training on the dataset EmotioNet using the model convnext_base:

python run.py --dataset EmotioNet --model convnext_base -bs 128 --fair

Output

The output will be saved to Fair-Consistent-Affect-Analysis/output

• checkpoint
  • *.pth: Checkpoint file
  • *_fair_results.pkl: Detailed fairness results
  • *_inference_results.pkl: Inference results of the test set
• tensorboard: TensorBoard file
• best_result.txt: Best result of global validation and its fairness result
• fair_result: Fairness results of each epoch
• result: Global validation results of each epoch

output example:

DISFA-convnext_base-[0524-1725]-lr0.0001-bs128                   
├── checkpoint                  
│   ├── convnext_base_DISFA_best_f1_macro_threshold0.8.pth
│   ├── convnext_base_DISFA_best_f1_macro_threshold0.8_fair_results.pkl
│   └── convnext_base_DISFA_best_f1_macro_threshold0.8_inference_results.pkl
├── tensorboard
│   └── events.out.tfevents.abc.43074.0
├── best_result.txt
├── fair_result.txt
├── loss.txt
└── result.txt

9 Testing

• eval: Activate Evaluation mode.
• checkpoint_path: Path to the checkpoint pth file

Training Examples

• To test the vit_b_16 checkpoint pre-trained on the dataset EmotioNet:

python run.py --dataset AffectNet-VA -bs 128 --fair --eval --ckeckpoint_path /path/to/vit_b_16_AffectNet-VA.pth

• To test the resnext50 checkpoint pre-trained on the dataset RAF-DB:

python run.py --dataset AffectNet-7 -bs 128 --fair --eval --ckeckpoint_path /path/to/resnext50_32x4d_RAF-DB.pth

10 Acknowledgements

This repo is based on the following projects, We thank the authors a lot for their valuable efforts.

POSTER++, DAN, HSEmotion, MA-Net, EAC, DACL, Emotion-GCN, ME-GraphAU, AUNets

11 Citation

Please cite the below paper in your publications if you use our developments (annotations, partitions, and/or trained models). BibTeX reference is as follows:

@article{hu2024bridging,
  title={Bridging the Gap: Protocol Towards Fair and Consistent Affect Analysis},
  author={Hu, Guanyu and Papadopoulou, Eleni and Kollias, Dimitrios and Tzouveli, Paraskevi and Wei, Jie and Yang, Xinyu},
  journal={arXiv preprint arXiv:2405.06841},
  year={2024}
}