BlEmoRe competition dataset

News

2025-12-23

New website launched: https://www.blemore.tech/

2025-12-10

Test dataset released, see the file test.zip. Please provide your predictions in the submission_template.json format (included in this Zenodo record). This file contains the required structure for submissions. Please email your predictions to Petri Laukka, and we will get back to you shortly with your accuracy scores. 

Access Instructions

This dataset is under restricted access and is only available to participants of the BlEmoRe Competition.

To gain access, either use the Zenodo portal to request access or send a short email to Petri Laukka
Please include the following:

• A statement that you agree to the BlEmoRe Dataset Limited Use License.

• Confirmation that you intend to participate in the competition.

• Your full name and affiliation, such as your university. 

Dataset Description

The dataset includes a total of 3050 multimodal portrayals from a total of 58 actors and actresses of both basic and blended emotions recorded under controlled studio conditions with synchronized audio and video.

• Training partition (train_01.zip - train_05.zip) includes 2456 portrayals from 47 actors and actresses.
• Test partition (test.zip) includes 594 portrayals from a total of 11 actors and actresses.

Each actor or actress was instructed to express:

Basic Emotions

• Anger, Disgust, Fear, Happiness, Sadness — portrayed in four different intensities (int1–int4)

• Neutral — portrayed in four distinct but equivalent imagined situations (sit1–sit4). These situations are not ordered or ranked, and are included primarily to increase variability in neutral portrayals

Blended Emotions

• All pairwise combinations of the five basic emotions (excluding neutral)

• Portrayed in three blend ratios:

  • 50/50: equal salience

  • 70/30: first emotion dominant

  • 30/70: second emotion dominant

Expressions were conveyed using facial expressions, body movement, and non-linguistic vocalizations (e.g., sighs, laughs). Participants were instructed to appear naturalistic and avoid exaggerated or stereotypical portrayals.

File Format & Naming Convention

The dataset is split into zip archives:

• 5 archives (train_01.zip - train_05.zip) with raw train .mov video files.
• 1 archive (test.zip) with raw test .mov video files.
• 1 archive (pre_extracted_train.zip) with pre-extracted features for the training files. 
• 1 archive (pre_extracted_test.zip) with pre-extracted features for the test files. 

Inside the train archives, each file follows the naming pattern:

Basic emotions

Blended emotions

• <video_id>: unique identifier for the actor/actress

• <emotion>: one of the five basic emotions

• <intensity_or_situation>: int1–int4 for intensity, sit1–sit4 for neutral context

• <version>: ver1, ver2, etc. (some portrayals have multiple valid takes)

• mix: appears only in blended emotion filenames

Metadata

A metadata file for the training partition is provided, containing:

• Filename

• Emotion(s) and blend ratio

• Actor/actress gender

• Suggested validation fold assignment — each fold contains non-overlapping actors/actresses, and the folds are generally balanced by gender.