MAFW

Details

MAFW is a large-scale, multi-modal, compound affective database for dynamic facial expression recognition in the wild. Clips in this database come from China, Japan, Korea, Europe, America and India, and cover various themes, e.g., variety, family, science fiction, suspense, love, comedy, and interviews, encompassing a wide range of human emotions. Each clip has been independently labeled 11 times by 11 well-trained annotators. MAFW database has enormous diversities, large quantities, and rich annotations, including:

Sample Video-audio Clips

1. Examples of the single expressions in MAFW

AngerEnglish: A girl with tears in her eyes shouts at the person opposite her. The deep frown,a downward pull on the lip corners,the higher inner corners of eyebrows and the lower outer corners of eyebrows.
中文：一个女生眼含着泪水大声训斥着对面的人。眉头紧蹙，嘴角下拉，眉毛内高外低。
DisgustEnglish: A woman looks nervously at her feet. The frown,the closed eyes and the open mouth.
中文：一个女人紧张的看着脚下的东西。皱眉，眼睛微闭，嘴巴张开。
FearEnglish: A girl gasps in the dark. The wide eyes and the open mouth.
中文：一个女孩在昏暗的环境中急促的喘息。瞪眼，嘴巴张大。
HappinessEnglish: A woman communicates with a man, talking about dinner. The slightly closed eyes, the open mouth and the raised lip corners.
中文：一个女人与男人交流，谈论着晚餐。眼睛微闭，嘴巴张开，嘴角上扬。
SadnessEnglish: A girl stands on the beach, tilting her head back and crying. The deep frown and the wide open mouth.
中文：一个女孩站在海边，仰着头哭泣。眉头紧蹙，嘴巴张大。
SurpriseEnglish: The woman sits with her eyes slowly widening, then suddenly shivers back and asks in a confused voice. The wide eyes.
中文：女人坐着眼睛慢慢睁大，然后突然向后颤动了一下，慌乱地询问道。瞪眼。
ContemptEnglish: A woman gives a dismissive laugh. A curl of the lips.
中文：一个女人不屑地笑了一声。撇嘴。
AnxietyEnglish: A woman can't get through to the phone and grumbles impatiently. The frown and the compressed lips.
中文：一个女人打不通电话，不耐烦地埋怨了一声。皱眉，抿嘴。
HelplessnessEnglish: A maid stands thoughtfully behind her mistress. The wide eyes and the compressed lips.
中文：一个侍女若有所思地站在女主人身后。瞪眼，抿嘴。
DisappointmentEnglish: A man lowers his head and takes a deep breath. The open mouth and the trembling lips.
中文：一个男人低下头，深呼吸。嘴巴半张，嘴唇颤抖。
Neutral-

2. Examples of the multiple expressions in MAFW

Anger
Disgust		English: A man glances his head to the side at the words of the man in front of him. The frown and a downward pull on the lip corners.
中文：一个男人听到面前的男人的话语后将头瞥向一边。皱眉，嘴角下拉。
Fear
Sadness		English: A woman squats on the ground and talks to a man tearfully. The marked frown and a downward pull on the lip corners.
中文：一个女人蹲坐在地上满含泪水的在和男人说话。皱眉，嘴角下拉。
Fear
Sadness
Anxiety		English: A woman begs someone. The marked frown, the wide eyes and the slightly open mouth.
中文：一个女人哀求着。皱眉，瞪眼，嘴巴微张。
Anger
Disgust
Contempt		English: A woman makes a condescending sarcasm at the other person. The raised chin, the raised eyebrows, the closed eyes and the frown.
中文：一个女人居高临下地挖苦对方。下巴上扬，挑眉，闭眼，皱眉。

Terms & Conditions

How to get the MAFW Dataset

This database is publicly available and free for professors and research scientists affiliated to a university. For students interested in accessing the dataset, please note that the application requires formal endorsement by a faculty member from your institution.

Permission to use (but not reproduce or distribute) the MAFW database is granted only if the following steps are properly followed:

  1. Download the MAFW-academics -final.pdf document, which serves as the End-User License Agreement (EULA).
  2. Carefully review the terms and conditions to confirm acceptance. The required information at the end of the document must be completed and signed—for student applicants, this signature must be from a professor at their affiliated university to validate the request.
  3. Send the fully completed and signed document to: 1202411179@cug.edu.cn.
  4. After review and approval, you will receive download links via email, including two options: Baidu Netdisk and Google Drive.

Citation

Please cite our paper if you find our work useful for your research:

@inbook{liu_mafw_2022,
	author = {Liu, Yuanyuan and Dai, Wei and Feng, Chuanxu and Wang, Wenbin and Yin, Guanghao and Zeng, Jiabei and Shan, Shiguang},
	title = {MAFW: A Large-scale, Multi-modal, Compound Affective Database for Dynamic Facial Expression Recognition in the Wild},
	year = {2022}
	isbn = {978-1-4503-9203-7},
	publisher = {ACM},
	address = {New York, NY, USA},
	url = {https://doi.org/10.1145/3503161.3548190},
	booktitle = {Proceedings of the 30th ACM International Conference on Multimedia (MM’22)},
	numpages = {9}
}

Content Preview

For more details of the dataset, please refer to the paper: MAFW: A Large-scale, Multi-modal, Compound Affective Database for Dynamic Facial Expression Recognition in the Wild.

For more details of emotional descriptive texts, please refer to supplementary materials for MAFW.

Frequently Asked Questions (FAQ)

After your application is approved, you will receive two download options:

2. What should I do if I can’t access Baidu Netdisk but need frame data?

If you require frame data but cannot access Baidu Netdisk, you can extract and process frames from the video clips obtained via Google Drive using the following steps (consistent with the preprocessing pipeline in our paper):

  1. Extract frame pictures from video clips using tools like OpenCV.
  2. Face detection and landmark extraction: Use face recognition tools (e.g., face-alignment-master referenced in our paper, or other convenient face detection libraries) to identify face regions and 68 facial landmarks.
  3. Face alignment and resizing: Perform affine transformation and matrix rotation (via OpenCV or similar libraries) to align faces, then resize the aligned face regions to 224×224 pixels (consistent with the dataset’s standard format).

You may use any familiar face detection, landmark extraction, or alignment tool that suits your workflow— the key is to ensure the final output is 224×224 aligned face frames for consistency with the dataset’s benchmark settings.

3. How to handle split compressed files (e.g., clips and frames)?

Due to the large size of the clips and frames directories, they are split into multiple compressed files:

Extraction instructions:

Critical notes:

  1. Ensure all split files are in the same folder (do not separate them into subdirectories).
  2. Do not rename any split files (e.g., avoid changing clips.7z.001 to clips_part1.7z), as this will break the extraction sequence.
  3. Verify that all files are fully downloaded (no corruption or missing parts) – incomplete downloads will cause extraction failures.
  4. The final extracted folder will be named clips or frames (no need to manually merge folders).

Code

The source code of our proposed T-ESFL model can be downloaded in https://github.com/MAFW-database/MAFW.

Contact

Please contact us for any questions about MAFW.

Yuanyuan LiuAssociate Professor, China University of Geosciencesliuyy@cug.edu.cn
Shuyang LiuMaster, China University of Geosciences20171003670@cug.edu.cn
Ying QianMaster, China University of Geosciences1202411179@cug.edu.cn

For more information, welcome to visit our team’s homepage: https://cvlab-liuyuanyuan.github.io/