FER2013

FER2013 is a benchmark dataset of 35,887 grayscale facial images used to train and evaluate emotion recognition models across seven distinct categories.

Pierre-Luc Carrier and Aaron Courville designed this dataset for the ICML 2013 Challenges in Representation Learning. It contains 48x48 pixel images labeled with seven emotions: anger, disgust, fear, happiness, sadness, surprise, and neutral. While the dataset is known for its high noise levels and label ambiguity, it remains a critical baseline for testing convolutional neural networks (CNNs) in real-time facial expression analysis.

https://www.kaggle.com/datasets/msambare/fer2013