Fer Gimefive Eval

---
name: fer-gimefive-eval
description: Evaluates the ability of convolutional neural networks to classify facial expressions into six basic emotion categories (happiness, surprise, sadness, anger, disgust, fear) from static images and video frames. The benchmark tests both classification accuracy and the model's capacity to generalize across in-the-wild and posed facial expression datasets. Use when the user wants to benchmark on RAF-DB, FER2013, FER GiMeFive, or asks about evaluating this task. Reports Accuracy (%).
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2402.15662
  bibtex_key: wang2024gimefive
  confidence: high
---

# fer-gimefive-eval

> GiMeFive: Towards Interpretable Facial Emotion Classification — Wang et al. (2024) (arXiv:2402.15662, 2024)

## What this evaluates

Evaluates the ability of convolutional neural networks to classify facial expressions into six basic emotion categories (happiness, surprise, sadness, anger, disgust, fear) from static images and video frames. The benchmark tests both classification accuracy and the model's capacity to generalize across in-the-wild and posed facial expression datasets.

## Datasets

- **RAF-DB** — total 12135; splits: train (9747), test (2388)
- **FER2013** — total 29688; splits: train (23743), test (5945)
- **FER GiMeFive** — total 63406; splits: train (55073), test (8333), valid (600); repo https://github.com/werywjw/data

## Metrics

- `Accuracy (%)` **(primary)** — range: percent
  - Standard classification accuracy: the proportion of correctly predicted emotion labels out of the total number of samples, reported as a percentage.

## Input / output format

**Input**: 64x64 pixel images converted to grayscale but stored as 3-channel tensors, normalized. For video evaluation, frames are face-cropped via Viola-Jones detection, resized to 64x64, converted to 3-channel grayscale, and normalized.

**Output**: Softmax probability distribution over six emotion classes (0: happiness, 1: surprise, 2: sadness, 3: anger, 4: disgust, 5: fear).

## Scoring recipe

```python
def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    return (correct / len(gold_labels)) * 100
```

## Common pitfalls

- The dataset uses grayscale images but pads them to 3 channels to satisfy CNN input dimensions, which may cause confusion if evaluators expect standard RGB inputs.
- The custom 'FER GiMeFive' dataset aggregates training and test splits from multiple heterogeneous benchmarks (RAF-DB, FER2013, etc.), so standard benchmark splits do not apply directly.
- Video evaluation relies on real-time frame extraction and Viola-Jones face detection, introducing preprocessing variability not captured in the static image accuracy tables.

## Evidence (verbatim from paper)

> Table 2: Accuracies (%) for different models with specific architectures and numbers of parameters and layers in our experiments (Note that ⋆ denotes our best GiMeFive; BN stands for the batch normalization, RB for residual block, SE for the squeeze and excitation block, and DO for dropout; +/- represent with/without respectively).

## Citation

```bibtex
@misc{wang2024gimefive,
  title={GiMeFive: Towards Interpretable Facial Emotion Classification},
  author={Wang et al. (2024)},
  year={2024},
  note={arXiv:2402.15662}
}
```

- arXiv: 2402.15662