DEEMO contains de-identified video clips of real people. Access is granted for non-commercial academic research only, under the DEEMO Data Use Agreement (full text in the dataset card below). Please use your institutional email address. Requests without a clear research purpose or institutional affiliation may be declined.

DEEMO: De-identity Multimodal Emotion Recognition and Reasoning

DEEMO is a dataset for emotion recognition and reasoning from de-identified video and audio. Faces are blurred and voices are de-identified, so models must rely on identity-free cues such as non-facial body language, speech content, vocal cues, and context rather than facial expressions or speaker identity.

The clips come from post-match press conferences and interviews of professional athletes, mostly tennis players, along with athletes from combat sports and other sports. Each clip has a binary emotion label, a transcript, a sequence of non-facial body language (NFBL) labels, textual descriptions of the audio and visual cues, and a free-text emotion reasoning.

Access

This dataset is gated with manual review. To request access:

Log in to Hugging Face and fill in the form at the top of this page. Use your institutional email address.
Read the Data Use Agreement below and sign it by typing your full name.
The dataset authors review each request. Once approved, you can download the data with your own Hugging Face account.

Access is personal. Every person who works with the data needs their own approved account.

Dataset structure

train/   1,229 clips (.mp4) + metadata.jsonl
test/      831 clips (.mp4) + metadata.jsonl

Each line of metadata.jsonl describes one clip:

Field	Type	Description
`file_name`	string	Clip file name: `{video_id}._{start_time}_{end_time}.mp4`
`video_id`	string	ID of the source video the clip was cut from
`start_time`, `end_time`	float	Clip position in the source video (seconds)
`conversation`	string	Transcript of the speech in the clip
`non_facial_body_languages`	list[string]	NFBL labels in order of appearance. A label can repeat when the behaviour occurs more than once.
`audio_cues`	string	Description of vocal cues (pace, pitch, intensity, rhythm)
`video_cues`	string	Description of the visual scene and the speaker's behaviour
`emotion`	string	`positive` or `negative`
`language`	string	Spoken language (all `English`)
`gpt_reasoning`	string	Free-text explanation of the emotion label, grounded in the cues above (field `GPT` in the original annotation files)

Example (long text fields shortened):

{
  "file_name": "627_34_0._28.8_37.7.mp4",
  "video_id": "627_34_0",
  "start_time": 28.8,
  "end_time": 37.7,
  "conversation": "She makes you really be on every single ball. I struggled a little bit. ...",
  "non_facial_body_languages": ["scratching or touching facial parts"],
  "audio_cues": "The speech in the audio is delivered at a moderate pace, exhibits a steady pitch ...",
  "video_cues": "The video shows a person seated at a table with microphones ...",
  "emotion": "negative",
  "language": "English",
  "gpt_reasoning": "The athlete's overall emotional state seems to lean towards the negative. ..."
}

Statistics

	Train	Test	Total
Clips	1,229	831	2,060
Source videos	294	195	489
Positive	757 (61.6%)	479 (57.6%)	1,236 (60.0%)
Negative	472 (38.4%)	352 (42.4%)	824 (40.0%)
Mean clip duration	11.8 s	10.9 s	11.4 s
Clip duration range	0.14–149.6 s	0.34–81.1 s	0.14–149.6 s
Total duration	4.02 h	2.52 h	6.53 h

Train and test clips come from disjoint sets of source videos.

Usage

Once your access request is approved, log in with your own account (hf auth login), then:

from datasets import load_dataset

ds = load_dataset("Jasonss/DEEMO")
print(ds["train"][0]["emotion"])

To download the raw files instead:

from huggingface_hub import snapshot_download

path = snapshot_download("Jasonss/DEEMO", repo_type="dataset")

Data Use Agreement

By requesting access and signing with your full name, you agree to the following terms with the DEEMO dataset authors. Access is personal: everyone who uses the data needs their own approved Hugging Face account.

Research only. Use the dataset for non-commercial academic research and teaching only. You may publish results, and models trained on the dataset for non-commercial research, but not the data itself.
No redistribution. Do not share or upload the dataset, or any part of it, anywhere.
No re-identification. Do not try to identify anyone in the dataset or reverse its de-identification. Do not create synthetic media of them, such as deepfakes or voice clones.
Publications. Show only a few de-identified frames for illustration, and cite the DEEMO paper.
Security and deletion. Store the data securely. Delete all copies when your access is revoked, when your research ends, or when we ask you to remove specific clips.
Copyright. The source footage belongs to its rights holders. You are responsible for complying with applicable law.
Reporting. Report any misuse or unauthorized disclosure to leedengsh@gmail.com within 7 days.
No warranty. The dataset is provided "as is", and access may be revoked at any time.

Citation

If you use DEEMO, please cite:

@inproceedings{li2025deemo,
  title     = {DEEMO: De-identity Multimodal Emotion Recognition and Reasoning},
  author    = {Li, Deng and Xing, Bohao and Liu, Xin and Xia, Baiqiang and Wen, Bihan and K{\"a}lvi{\"a}inen, Heikki},
  booktitle = {Proceedings of the 33rd ACM International Conference on Multimedia},
  pages     = {5707--5716},
  year      = {2025},
  doi       = {10.1145/3746027.3755411}
}

Contact

For questions, removal requests, or to report misuse, contact leedengsh@gmail.com.

Downloads last month: 15

Total file size:

4.7 GB

Paper for Jasonss/DEEMO

DEEMO: De-identity Multimodal Emotion Recognition and Reasoning

Paper • 2504.19549 • Published Oct 25, 2025