File size: 1,503 Bytes
3911a78
 
5ceb609
 
 
 
 
3911a78
5ceb609
053bfbe
5ceb609
053bfbe
5ceb609
053bfbe
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
---
license: mit
task_categories:
- object-detection
- image-classification
tags:
- medical
---

Face Masks ensemble dataset is no longer limited to [Kaggle](https://www.kaggle.com/datasets/henrylydecker/face-masks), it is now coming to Huggingface!

This dataset was created to help train and/or fine tune models for detecting masked and un-masked faces.

I created a new face masks object detection dataset by compositing together three publically available face masks object detection datasets on Kaggle that used the YOLO annotation format. 
To combine the datasets, I used Roboflow. 
All three original datasets had different class dictionaries, so I recoded the classes into two classes: "Mask" and "No Mask". 
One dataset included a class for incorrectly worn face masks, images with this class were removed from the dataset. 
Approximately 50 images had corrupted annotations, so they were manually re-annotated in the Roboflow platform. 
The final dataset includes 9,982 images, with 24,975 annotated instances. 
Image resolution was on average 0.49 mp, with a median size of 750 x 600 pixels.

To improve model performance on out of sample data, I used 90 degree rotational augmentation. 
This saved duplicate versions of each image for 90, 180, and 270 degree rotations. 
I then split the data into 85% training, 10% validation, and 5% testing. 
Images with classes that were removed from the dataset were removed, leaving 16,000 images in training, 1,900 in validation, and 1,000 in testing.