Upload README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,164 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
pretty_name: MLLM Visual Emotion
|
| 4 |
+
task_categories:
|
| 5 |
+
- image-classification
|
| 6 |
+
- visual-question-answering
|
| 7 |
+
language:
|
| 8 |
+
- en
|
| 9 |
+
tags:
|
| 10 |
+
- emotion-recognition
|
| 11 |
+
- facial-expression
|
| 12 |
+
- multimodal
|
| 13 |
+
- large-language-models
|
| 14 |
+
- benchmark
|
| 15 |
+
- visual-emotion-analysis
|
| 16 |
+
size_categories:
|
| 17 |
+
- 1K-10K
|
| 18 |
+
viewer: false
|
| 19 |
+
configs:
|
| 20 |
+
- config_name: default
|
| 21 |
+
data_files: []
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
# MLLMVisualEmotion
|
| 25 |
+
|
| 26 |
+
This repository accompanies the study 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.
|
| 27 |
+
|
| 28 |
+
## Repository Structure
|
| 29 |
+
|
| 30 |
+
```
|
| 31 |
+
MLLMVisualEmotion/
|
| 32 |
+
├── datasets/
|
| 33 |
+
│ ├── Emotion6/
|
| 34 |
+
│ │ └── processed/
|
| 35 |
+
│ │ ├── emotion_dataset.json
|
| 36 |
+
│ │ ├── images_multiple_choice_questions.json
|
| 37 |
+
│ │ ├── images_judgement_questions.json
|
| 38 |
+
│ │ ├── images_marked_multiple_choice_questions.json
|
| 39 |
+
│ │ ├── images_could_judgement_questions.json
|
| 40 |
+
│ │ ├── images_really_judgement_questions.json
|
| 41 |
+
│ │ ├── images_emotional_tendency_multiple_choice_questions.json
|
| 42 |
+
│ │ ├── images_emotional_tendency_judgement_questions.json
|
| 43 |
+
│ │ └── imgs/
|
| 44 |
+
│ └── FaceEmotion/
|
| 45 |
+
│ └── processed/
|
| 46 |
+
│ ├── emotion_dataset.json
|
| 47 |
+
│ ├── images_multiple_choice_questions.json
|
| 48 |
+
│ ├── images_judgement_questions.json
|
| 49 |
+
│ ├── images_marked_multiple_choice_questions.json
|
| 50 |
+
│ ├── images_could_judgement_questions.json
|
| 51 |
+
│ ├── images_really_judgement_questions.json
|
| 52 |
+
│ ├── images_emotional_tendency_multiple_choice_questions.json
|
| 53 |
+
│ ├── images_emotional_tendency_judgement_questions.json
|
| 54 |
+
│ └── imgs/
|
| 55 |
+
├── model_outputs/
|
| 56 |
+
│ ├── Undeterminable_results.json
|
| 57 |
+
│ ├── Gemini3p1Flash/
|
| 58 |
+
│ ├── GPT4o/
|
| 59 |
+
│ ├── GPT5/
|
| 60 |
+
│ ├── InternVL2/
|
| 61 |
+
│ ├── InternVL3/
|
| 62 |
+
│ ├── LLaVA1p5/
|
| 63 |
+
│ ├── LLaVA1p6/
|
| 64 |
+
│ ├── Qwen2/
|
| 65 |
+
│ ├── Qwen2p5/
|
| 66 |
+
│ └── Qwen3/
|
| 67 |
+
└── Source_Data.xlsx
|
| 68 |
+
```
|
| 69 |
+
|
| 70 |
+
## Datasets
|
| 71 |
+
|
| 72 |
+
### Emotion6 (Preprocessed)
|
| 73 |
+
|
| 74 |
+
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.
|
| 75 |
+
|
| 76 |
+
- **`emotion_dataset.json`** — Image filenames, emotion labels, and arousal values.
|
| 77 |
+
- **`imgs/`** — The image files.
|
| 78 |
+
|
| 79 |
+
### FaceEmotion (Newly Constructed)
|
| 80 |
+
|
| 81 |
+
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*).
|
| 82 |
+
|
| 83 |
+
- **`emotion_dataset.json`** — Image filenames and emotion labels.
|
| 84 |
+
- **`imgs/`** — The image files.
|
| 85 |
+
|
| 86 |
+
### Task Input Files
|
| 87 |
+
|
| 88 |
+
For each dataset, the following JSON files contain the formatted prompts used as inputs to the MLLMs:
|
| 89 |
+
|
| 90 |
+
| File | Task Description |
|
| 91 |
+
|------|-----------------|
|
| 92 |
+
| `images_multiple_choice_questions.json` | Multi-class emotion selection (7-way classification) |
|
| 93 |
+
| `images_judgement_questions.json` | Binary emotion judgement (Yes/No per emotion category) |
|
| 94 |
+
| `images_marked_multiple_choice_questions.json` | Multi-class emotion selection with contextual cues |
|
| 95 |
+
| `images_could_judgement_questions.json` | Binary emotion judgement with "Could" contextual framing |
|
| 96 |
+
| `images_really_judgement_questions.json` | Binary emotion judgement with "really" contextual framing |
|
| 97 |
+
| `images_emotional_tendency_multiple_choice_questions.json` | Multi-class emotion selection with emotional tendency framing |
|
| 98 |
+
| `images_emotional_tendency_judgement_questions.json` | Binary emotion judgement with emotional tendency framing |
|
| 99 |
+
|
| 100 |
+
Each entry in these files contains:
|
| 101 |
+
- **Image identifier** and path
|
| 102 |
+
- **Ground-truth label**
|
| 103 |
+
- **Experiment setup** identifier
|
| 104 |
+
- **Query** (the full prompt sent to the model)
|
| 105 |
+
- **Answer** (the ground-truth answer)
|
| 106 |
+
- **Arousal** value (Emotion6 only)
|
| 107 |
+
|
| 108 |
+
## Model Outputs
|
| 109 |
+
|
| 110 |
+
The `model_outputs/` directory contains the raw outputs from 10 MLLMs evaluated on all tasks:
|
| 111 |
+
|
| 112 |
+
| Model | Directory |
|
| 113 |
+
|-------|-----------|
|
| 114 |
+
| Gemini 3.1 Flash | `Gemini3p1Flash/` |
|
| 115 |
+
| GPT-4o | `GPT4o/` |
|
| 116 |
+
| GPT-5 | `GPT5/` |
|
| 117 |
+
| InternVL2 | `InternVL2/` |
|
| 118 |
+
| InternVL3 | `InternVL3/` |
|
| 119 |
+
| LLaVA 1.5 | `LLaVA1p5/` |
|
| 120 |
+
| LLaVA 1.6 | `LLaVA1p6/` |
|
| 121 |
+
| Qwen2-VL | `Qwen2/` |
|
| 122 |
+
| Qwen2.5-VL | `Qwen2p5/` |
|
| 123 |
+
| Qwen3-VL | `Qwen3/` |
|
| 124 |
+
|
| 125 |
+
Each model directory contains two subdirectories:
|
| 126 |
+
- **`Emotion6/`** — Results on the natural-scene Emotion6 dataset.
|
| 127 |
+
- **`FaceEmotion/`** — Results on the facial-expression FaceEmotion dataset.
|
| 128 |
+
|
| 129 |
+
Within each subdirectory, result files mirror the task input files:
|
| 130 |
+
|
| 131 |
+
| File | Task Description |
|
| 132 |
+
|------|-----------------|
|
| 133 |
+
| `images_multiple_choice_results.json` | Multi-class emotion selection results |
|
| 134 |
+
| `images_judgement_results.json` | Binary emotion judgement results |
|
| 135 |
+
| `images_marked_multiple_choice_results.json` | Multi-class selection with contextual cues results |
|
| 136 |
+
| `images_could_judgement_results.json` | Binary judgement with "Could" framing results |
|
| 137 |
+
| `images_really_judgement_results.json` | Binary judgement with "really" framing results |
|
| 138 |
+
| `images_emotional_tendency_multiple_choice_results.json` | Multi-class selection with emotional tendency results |
|
| 139 |
+
| `images_emotional_tendency_judgement_results.json` | Binary judgement with emotional tendency results |
|
| 140 |
+
|
| 141 |
+
Each result entry extends the input entry with:
|
| 142 |
+
- **`AI_answer`** — The model's predicted answer.
|
| 143 |
+
- **`accuracy`** — Binary accuracy score (1 = correct, 0 = incorrect).
|
| 144 |
+
- **`model_output`** — The full raw text output from the model.
|
| 145 |
+
|
| 146 |
+
### Undeterminable Results
|
| 147 |
+
|
| 148 |
+
**`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.
|
| 149 |
+
|
| 150 |
+
## Source Data
|
| 151 |
+
|
| 152 |
+
**`Source_Data.xlsx`** contains the outcome tables and source data underlying all figures presented in the study.
|
| 153 |
+
|
| 154 |
+
## License
|
| 155 |
+
|
| 156 |
+
This dataset is released under the [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/) license.
|
| 157 |
+
|
| 158 |
+
## Citation
|
| 159 |
+
|
| 160 |
+
If you use this dataset in your research, please cite the corresponding paper.
|
| 161 |
+
|
| 162 |
+
## Contact
|
| 163 |
+
|
| 164 |
+
For questions or issues, please open a discussion on the Hugging Face repository.
|