hoangdang004/BoEM

Emotion Dataset

Dataset Description

Dataset Summary

BoEM: A Multimodal VQA Dataset for Fine-Grained Facial Emotion Recognition.

This dataset contains facial expressions with corresponding emotion labels. Each image is labeled with emotions such as fear, happiness, sadness, etc., along with intensity information.

Languages

The dataset annotations are in English.

Dataset Structure

Data Instances

Each instance in the dataset consists of:

• id: Unique identifier for the image
• image: The facial expression image
• instructions: The prompt/instruction given to annotate the image
• answer: The emotion label assigned to the image

Data Fields

• id: String - Unique identifier for each sample
• image: Image - The facial expression image
• instructions: String - The prompt/instruction given to annotate the image
• answer: String - The emotion label/description

Data Splits

The dataset is divided into training and validation splits.

Dataset Creation

Source Data

The dataset contains facial expressions collected for emotion recognition tasks.

Annotations

The images are annotated with emotion labels and intensity information.

Considerations for Using the Data

Licensing Information

MIT

@misc{emotionbench2026submission, title = {BoEM: A Multimodal VQA Dataset for Emotion Recognition}, author = {Hoang Nam Dang, Lockheed Martin, Mei-Huei Tang}, year = {2026}, note = {Manuscript submitted to CVPR 2026}, }