---
task_categories:
- image-classification
tags:
- emotions
size_categories:
- 100K<n<1M
---
A merged emotions dataset was created using a highly curated subset of ExpW, FER2013 (enhanced with FER2013+), AffectNet (6 emotions), and RAF-DB in YOLO format, totaling approximately 155K samples. A YOLOv11-x model, fine-tuned on the WiderFace dataset for the bounding boxes, was used. The distribution is as follows:

TRAIN Set Class Distribution:
--------------------
* Class 0 (Angry): 8511 (6.84%)
* Class 1 (Disgust): 6307 (5.07%)
* Class 2 (Fear): 4249 (3.41%)
* Class 3 (Happy): 37714 (30.30%)
* Class 4 (Neutral): 39297 (31.57%)
* Class 5 (Sad): 15809 (12.70%)
* Class 6 (Surprise): 12593 (10.12%)

VAL Set Class Distribution:
--------------------
* Class 0 (Angry): 1091 (6.98%)
* Class 1 (Disgust): 815 (5.21%)
* Class 2 (Fear): 583 (3.73%)
* Class 3 (Happy): 4722 (30.19%)
* Class 4 (Neutral): 4894 (31.29%)
* Class 5 (Sad): 1927 (12.32%)
* Class 6 (Surprise): 1607 (10.28%)

TEST Set Class Distribution:
--------------------
* Class 0 (Angry): 1080 (6.91%)
* Class 1 (Disgust): 771 (4.93%)
* Class 2 (Fear): 514 (3.29%)
* Class 3 (Happy): 4744 (30.36%)
* Class 4 (Neutral): 4962 (31.75%)
* Class 5 (Sad): 1998 (12.79%)
* Class 6 (Surprise): 1557 (9.96%)

TOTAL Set Class Distribution:
--------------------
* Class 0 (Angry): 10682 (6.86%)
* Class 1 (Disgust): 7893 (5.07%)
* Class 2 (Fear): 5346 (3.43%)
* Class 3 (Happy): 47180 (30.29%)
* Class 4 (Neutral): 49153 (31.56%)
* Class 5 (Sad): 19734 (12.67%)
* Class 6 (Surprise): 15,757 (10.12%)

As you can see, there is a significant imbalance between classes. We can see it more clearly here:

![Global distribution](https://huggingface.co/datasets/AdamCodd/yolo-emotions/resolve/main/metadata.jpg)

## Limitations

Many of the images do not represent the described emotions and some are merely meaningless grimaces. Moreover, some images appear multiple times across the train, validation, and test sets. Honestly, I wonder how researchers manage to work with these kinds of datasets, despite their widespread use. Regardless, I did my best to improve it without rebuilding it from the ground up. 

The 'contempt' emotion in AffectNet was also removed because it was not included in the other datasets and would have contributed to greater class imbalance.