This repository accompanies the study "Reading Human Emotion Through Machine Eyes" on visual emotion recognition by multimodal large language models (MLLMs). It contains the newly constructed FaceEmotion dataset, the preprocessed Emotion6 dataset, outcome tables, source data underlying all figures, and the complete model outputs for all tasks across all scenarios.

Repository Structure

MLLMVisualEmotion/
├── datasets/
│   ├── Emotion6/
│   │   └── processed/
│   │       ├── emotion_dataset.json
│   │       ├── images_multiple_choice_questions.json
│   │       ├── images_judgement_questions.json
│   │       ├── images_marked_multiple_choice_questions.json
│   │       ├── images_could_judgement_questions.json
│   │       ├── images_really_judgement_questions.json
│   │       ├── images_emotional_tendency_multiple_choice_questions.json
│   │       ├── images_emotional_tendency_judgement_questions.json
│   │       └── imgs/
│   └── FaceEmotion/
│       └── processed/
│           ├── emotion_dataset.json
│           ├── images_multiple_choice_questions.json
│           ├── images_judgement_questions.json
│           ├── images_marked_multiple_choice_questions.json
│           ├── images_could_judgement_questions.json
│           ├── images_really_judgement_questions.json
│           ├── images_emotional_tendency_multiple_choice_questions.json
│           ├── images_emotional_tendency_judgement_questions.json
│           └── imgs/
├── model_outputs/
│   ├── Undeterminable_results.json
│   ├── Gemini3p1Flash/
│   ├── GPT4o/
│   ├── GPT5/
│   ├── InternVL2/
│   ├── InternVL3/
│   ├── LLaVA1p5/
│   ├── LLaVA1p6/
│   ├── Qwen2/
│   ├── Qwen2p5/
│   └── Qwen3/
└── Source_Data.xlsx

Datasets

Emotion6 (Preprocessed)

The Emotion6 dataset is a widely used benchmark for visual emotion recognition, originally consisting of 1,980 images collected from Flickr. Each image is annotated with one of seven emotion categories (Anger, Disgust, Fear, Joy, Neutral, Sadness, Surprise) and an arousal value on a 1–9 scale.

emotion_dataset.json — Image filenames, emotion labels, and arousal values.
imgs/ — The image files.

FaceEmotion (Newly Constructed)

FaceEmotion is a newly constructed dataset of 1,000 facial expression images, each annotated with one of seven emotion categories (Anger, Disgust, Fear, Joy, Neutral, Sadness, Surprise).

emotion_dataset.json — Image filenames and emotion labels.
imgs/ — The image files.

Task Input Files

For each dataset, the following JSON files contain the formatted prompts used as inputs to the MLLMs:

File	Task Description
`images_multiple_choice_questions.json`	Multi-class emotion selection (7-way classification)
`images_judgement_questions.json`	Binary emotion judgement (Yes/No per emotion category)
`images_marked_multiple_choice_questions.json`	Multi-class emotion selection with contextual cues
`images_could_judgement_questions.json`	Binary emotion judgement with "Could" contextual framing
`images_really_judgement_questions.json`	Binary emotion judgement with "really" contextual framing
`images_emotional_tendency_multiple_choice_questions.json`	Multi-class emotion selection with emotional tendency framing
`images_emotional_tendency_judgement_questions.json`	Binary emotion judgement with emotional tendency framing

Each entry in these files contains:

Image identifier and path
Ground-truth label
Experiment setup identifier
Query (the full prompt sent to the model)
Answer (the ground-truth answer)
Arousal value (Emotion6 only)

Model Outputs

The model_outputs/ directory contains the raw outputs from 10 MLLMs evaluated on all tasks:

Model	Directory
Gemini 3.1 Flash	`Gemini3p1Flash/`
GPT-4o	`GPT4o/`
GPT-5	`GPT5/`
InternVL2	`InternVL2/`
InternVL3	`InternVL3/`
LLaVA 1.5	`LLaVA1p5/`
LLaVA 1.6	`LLaVA1p6/`
Qwen2-VL	`Qwen2/`
Qwen2.5-VL	`Qwen2p5/`
Qwen3-VL	`Qwen3/`

Each model directory contains two subdirectories:

Emotion6/ — Results on the natural-scene Emotion6 dataset.
FaceEmotion/ — Results on the facial-expression FaceEmotion dataset.

Within each subdirectory, result files mirror the task input files:

File	Task Description
`images_multiple_choice_results.json`	Multi-class emotion selection results
`images_judgement_results.json`	Binary emotion judgement results
`images_marked_multiple_choice_results.json`	Multi-class selection with contextual cues results
`images_could_judgement_results.json`	Binary judgement with "Could" framing results
`images_really_judgement_results.json`	Binary judgement with "really" framing results
`images_emotional_tendency_multiple_choice_results.json`	Multi-class selection with emotional tendency results
`images_emotional_tendency_judgement_results.json`	Binary judgement with emotional tendency results

Each result entry extends the input entry with:

AI_answer — The model's predicted answer.
accuracy — Binary accuracy score (1 = correct, 0 = incorrect).
model_output — The full raw text output from the model.

Undeterminable Results

Undeterminable_results.json records, for each model and each scenario (Emotion6 / FaceEmotion), the count and frequency of "Undeterminable" responses in the multi-class emotion selection task that includes an "Undeterminable" option.

Source Data

Source_Data.xlsx contains the outcome tables and source data underlying all figures presented in the study.

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Citation

If you use this dataset in your research, please cite the corresponding paper.

Contact

For questions or issues, please open a discussion on the Hugging Face repository.

Downloads last month: 68

Total file size:

1.42 GB