Bodily Expressed Emotion
Understanding (BEEU) Challenge

1. Multi-Label Emotion Prediction

Predict one or more emotion categories associated with each video clip.

Goal: Develop models capable of handling overlapping or blended emotions expressed through bodily cues.

Evaluation Criteria: Evaluation will use simple set-based metrics (intersection over union). (There will be bonus points even for misclassified samples, based on emotion distance). Evaluation will be performed only on the test set.

Required Response Format: The response submitted by the participant should include the video ID (for videos in the test set), and the list of predicted emotion words, where the set of emotions covers both the main categories and the sub-categories. This should ideally be in a CSV format, under the column names of id and prediction.

2. Emotional Explanation Generation

Generate textual explanations describing why a particular emotion is inferred to be expressed in the clip.

Goal: Encourage interpretability and reasoning in emotion recognition systems.

Evaluation Criteria: We will conduct curated human studies to evaluate the generated explanation. Further, we will use standard n-gram metrics like BLEU and ROUGE to evaluate the quality of explanations.

Required Response Format: For each label in the ground truth label set of the video, an explanation should be generated. The submission can be a JSON-like file, where the video ID is present, and emotion labels in the ground truth set serve as keys, and the generated explanations serve as values.

3. Valence–Arousal–Dominance (VAD) Prediction

Predict continuous valence, arousal, and dominance scores for each video.

Goal: Capture the underlying affective dimensions that complement categorical emotion representations.

Evaluation Criteria: We will use the R^2 metric for this regression task. Evaluation will be only on the test set.

Required Response Format: Similar to Task 1, the submission should include the video IDs (under column id), and along with that, include 3 other columns: valence, arousal, dominance with the predicted values of each.

4. Open-Ended Usage Track

In this track, participants are encouraged to submit any material (video, demo of created system) that demonstrates creative usage of our bodily expressed emotion dataset, beyond the requirements of the standard tasks listed above. This could include application of the dataset on any downstream task beyond emotion recognition, using it to create artistic or interactive systems, etc. The submission format will support upload of a video (for a demo or similar), and a PDF file detailing the method. Paper submissions on this are also highly encouraged (in which case the report can be omitted).