jihedjabnoun/faceemo-set-features

FaceEmo-Set Features (ViT Representations)

Dataset Illustration

Overview

This dataset provides feature representations extracted from multiple facial emotion recognition datasets using a Vision Transformer (ViT-Base/16) model.

Instead of releasing raw images, which cannot be redistributed because some source materials are subject to copyright restrictions, we provide:

• High-level feature embeddings
• Model outputs (logits and probabilities)
• Structured metadata for reproducibility

This enables reproducible research and downstream analysis without access to the original visual data.

Motivation

Traditional facial emotion recognition benchmarks often suffer from:

• Severe class imbalance (e.g., FER2013)
• Limited diversity of sources
• Poor cross-dataset generalization

This dataset is designed to support data-centric analysis and representation-level research, particularly in the context of:

• Cross-dataset evaluation
• Domain shift robustness
• Feature-space analysis
• Representation-level benchmarking
• Feature-space analysis of multimodal systems

Source Datasets

Feature representations are extracted from a combined dataset including:

• FaceEmo-Set
• FER2013
• RAF-DB
• RAVDESS
• AffectNet (used for evaluation only)

Each sample is associated with its original dataset source via metadata.

Feature Extraction

• Backbone: ViT-Base/16
• Pretraining: ImageNet-21k
• Fine-tuning: Combined dataset (FaceEmo-Set + FER2013 + RAF-DB + RAVDESS)
• Feature type: CLS token from the final hidden layer
• Feature dimension: 768

Features are extracted using the same model weights available at:

https://huggingface.co/jihedjabnoun/faceemo-set

Dataset Structure

Splits

Split	Samples
Train	62,902
Validation	15,946
Test (FER2013)	7,178
Test (AffectNet)	3,006

Feature Files

Each .npz file contains:

• features: shape (N, 768)
• logits: shape (N, 7)
• probs: shape (N, 7)

Files:

• combined_vit_features_train.npz
• combined_vit_features_val.npz
• combined_vit_features_test_fer2013.npz
• combined_vit_features_test_affectnet.npz

Metadata

combined_vit_metadata_all.csv

Columns:

• sample_id
• Dataset
• Emotion
• Split
• valid_image

Summary

combined_vit_feature_summary.csv

Labels

Seven emotion classes:

• anger
• disgust
• fear
• happiness
• neutral
• sadness
• surprise

🚀 How to Load the Dataset

Option 1 — Google Colab

from huggingface_hub import hf_hub_download
import numpy as np

train_path = hf_hub_download(
    repo_id="jihedjabnoun/faceemo-set-features",
    filename="combined_vit_features_train.npz",
    repo_type="dataset"
)

data = np.load(train_path)

X = data["features"]
logits = data["logits"]
probs = data["probs"]

print(X.shape)

Option 2 — Load Metadata

import pandas as pd
from huggingface_hub import hf_hub_download

csv_path = hf_hub_download(
    repo_id="jihedjabnoun/faceemo-set-features",
    filename="combined_vit_metadata_all.csv",
    repo_type="dataset"
)

df = pd.read_csv(csv_path)
print(df.head())

Option 3 — Download Full Dataset

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="jihedjabnoun/faceemo-set-features",
    repo_type="dataset"
)

print(local_dir)

Option 4 — CLI

huggingface-cli download jihedjabnoun/faceemo-set-features --repo-type dataset --local-dir ./faceemo_features

🔍 Feature Inversion & Reconstruction Analysis

To evaluate whether the released feature representations retain recoverable visual information, we conducted a feature inversion experiment using gradient-based optimization.

Method

• Input: CLS embeddings (768-dimensional)
• Model: the same ViT used for feature extraction
• Objective:
  • minimize feature distance in embedding space
  • optionally match logits
• Regularization:
  • total variation (TV)
  • L2 pixel prior

📊 Example Results

Original vs Reconstructed Image

Optimization Behavior

Quantitative Example

• Original predicted class: fear (0.72)
• Reconstructed predicted class: fear (0.71)
• Final cosine similarity: ~0.96
• Final CLS MSE: ~0.005

Key Observations

• High feature alignment was achieved in embedding space (cosine similarity > 0.95)
• Reconstructed images preserve:
  • coarse facial structure
  • emotion-related cues
  • semantic characteristics sufficient to recover the predicted class in this example
• Reconstructed images do not preserve:
  • identity
  • fine visual details
  • exact pixel-level appearance

Core Insight

High similarity in feature space does not imply faithful visual reconstruction.

Despite strong alignment in embedding space, the reconstructed image differs substantially from the original at the pixel level. This suggests that the released CLS representations retain semantic information while discarding most fine-grained spatial detail.

Conclusion

• CLS embeddings retain semantic information
• Exact image reconstruction is not observed
• Approximate inversion is possible but remains visually ambiguous

📓 Reproducible Notebook

The full inversion pipeline is provided in:

• upload_predict_cls_reconstruct_compare_colab.ipynb

This notebook allows users to:

• upload an image
• run prediction with the released model
• extract its CLS feature
• reconstruct an image from that feature
• compare original and reconstructed outputs

Applications

• Cross-dataset generalization studies
• Representation learning analysis
• Domain shift evaluation
• Multimodal fusion research
• Low-resource training

Limitations

• No raw images
• No temporal modeling
• Limited to 7 emotions
• Potential dataset bias
• No landmarks
• Feature inversion may reveal approximate semantic information

Ethical Considerations

Emotion recognition systems may:

• Encode biases
• Be misused in surveillance

Additionally:

• Feature representations may allow approximate inversion
• Reconstructed images in our experiments do not preserve identity or exact appearance

Users should:

• Evaluate fairness
• Avoid high-risk deployment
• Consider privacy implications of feature sharing

🔬 Reproducibility

All features were extracted using a single trained ViT model to ensure consistency.

The inversion example included in this repository is intended as a transparency and analysis tool to help users understand what information may remain in released embeddings.

Citation

@inproceedings{jabnoun2026faceemoset,
  title={Improving Cross-Dataset Generalization in Facial Emotion Recognition Through FaceEmo-Set: A Balanced and Diverse Dataset},
  author={Jabnoun, Jihed and Maraoui, Mohsen and Zrigui, Mounir},
  booktitle={18th Asian Conference on Intelligent Information and Database Systems (ACIIDS)},
  year={2026},
  address={Kaohsiung, Taiwan},
  month={April}
}

Acknowledgments

University of Monastir, Tunisia

License

MIT License