What Are Face Datasets?
Image datasets contain digital images designed to help train, test, and evaluate machine learning (ML) and deep learning (DL) algorithms used for computer vision systems. Managing and optimizing datasets is one of the key steps in the machine learning operations (MLOps) pipeline.
Face datasets are image datasets with face images curated for ML projects. Instead of collecting your own training data, you can use one of several publicly available face datasets.
Face datasets typically contain faces in different locations and lighting conditions, representing human emotions, ethnicity, age, and other characteristics.
Face datasets are a key building block of facial recognition technology. There are many use cases in this field of computer vision, including video surveillance, mobile authentication, and augmented reality (AR).
Common Use Cases for Face Datasets
Facial Recognition
This use case includes algorithms that perform face identification, recognition, and clustering tasks (grouping similar faces). The best algorithms use a combination of face preprocessing and face alignment to improve face recognition. These algorithms typically use a multi-task cascading convolutional network (MTCNN) to detect faces and find landmarks.
Facial landmark recognition can also be used to identify emotions. Emotions are typically detected by analyzing movement of the lips, eyes, and eyebrows.
Facial Motion Capture
Some popular apps like Snapchat and Instagram allow users to edit faces in real time with fun filters. This is achieved through a face recognition algorithm that tells the app that there is a face on the screen that can be tracked and corrected.
Face detection technology is also enabling creation of real-time avatars for computer graphics, 3D animation and movies, video games and other media. Facial emotion tracking is an important component of realistic animated avatars.
Because computer-generated character movements are extracted from real human gestures, the resulting animations look more natural and detailed than hand-made ones.
Driver Monitoring
Driver fatigue is the cause of many car accidents, and smart car emergency stop functions are not always effective. Monitoring drivers for signs of fatigue can help prevent accidents and save lives. For example, computer vision models can process video from in-vehicle cameras to detect signs of fatigue or distraction on the driver’s face. If the driver is not paying enough attention to the road, the model can trigger a warning.
Built-in motion tracking systems and wearable electrocardiogram (ECG) tracking devices can perform similar functions, but they must be worn on the human body, which is inconvenient for users. Computer vision models offer a simpler and less invasive approach. A neural network can learn to use facial landmark input to identify drowsiness on a driver’s face.
Another approach is the MobileNetV2 architecture, which allows the detection of driver fatigue in a video stream without the use of facial landmarks. The downside of this model is that it requires longer training time. Neural network inference speed and mobile device quality can affect driver monitoring systems. So in many cases, the landmark detection algorithm provides a better overall solution.
Top 4 Face Datasets
CelebA
CelebFaces Attributes Dataset (CelebA) is a face attributes dataset made available for research and non-commercial purposes only. This large-scale dataset provides over 200,000 images of celebrities that cover many pose variations and background clutter.
You can use CelebA to train and test sets for various computer vision tasks, including face attribute recognition, landmark localization, face editing and synthesis, and face detection. The dataset includes 202,599 face images, 10,177 identities, five landmark locations, and 40 binary attributes annotations in each image.
Flickr-Faces-HQ Dataset (FFHQ)
Flickr-Faces-HQ Dataset (FFHQ) provides 70,000 high-quality PNG images of human faces at 10241024 resolution. It contains various ethnicities, ages, and image backgrounds. Compared to CelevA, FFHQ offers more variety, including accessories like eyeglasses, hats, and sunglasses.
FFHQ was originally created as a benchmark for generative adversarial networks (GAN). The images were crawled from Flickr and automatically cropped and aligned.
300W
300-W is a face dataset that includes 300 images of an indoor setting and 300 images set outdoors. These images cover diverse identities, lighting conditions, facial expressions, poses, face sizes, and occlusions.
These images include annotations of 68 point markers applied with a semi-automatic method. All images were carefully selected to represent a challenging sample of natural face instances under unrestricted conditions.
AFLW
Annotated Facial Landmarks in the Wild (AFLW) provides annotated facial images that were sourced from Flickr. These images represent various poses, ethnicities, facial expressions, ages, genders, and imaging conditions. The dataset includes 25,000 face images, each annotated with up to 21 landmarks.
Tutorial: Working With the LFW Dataset
LFW (Labeled faces in the wild) is a library of face images created to examine the challenge of unconstrained face identification. The LFW dataset includes two loaders: fetch_lfw_people and fetch_lfw_pairs, that are used for face identification and face verification, respectively. This guide employs the memmap edition found in ~/scikit_learn_data/lfw_home/ by the mean of the joblib tool.
First, we will use and implement the fetch_lfw_people loader. This loader classifies faces into many classes using supervised learning. This tutorial demonstrates how to import the labeled faces in the wild dataset and display the name of the person present in an image.
Follow these step to use the fetch_lfw_people loader:
- First, import the loader from
a sklearn.datasetspackage.
from sklearn.datasets import fetch_lfw_people
2. We will use the fetch_lfw_people command, which will contain images of individuals with at least 70 distinct images. Run the command below to retrieve the dataset and loader:
lfw_people = fetch_lfw_people(min_faces_per_person=70, resize=0.4)
3. Run the following code to display the names of the individuals contained in the dataset:
for indiviual_name in lfw_people.target_names:
print(indiviual_name)
The dataset consists of pictures of people, and each one has a unique identifier from the target array associated with it.
4. The code below will allow the user to retrieve the ground truth data by the target array:
lfw_people.target.shape
list(lfw_people.target[:10])
Lastly, we will use and implement the fetch_lfw_pairs loader. The loader is useful for determining whether or not two images depict the same person. When retrieving the loader, it is crucial to specify the specific subset of the dataset.
Follow these step to use the fetch_lfw_pairs loader:
- First, import the loader from a sklearn.datasets package.
from sklearn.datasets import fetch_lfw_pairs
2. After importing the loader, users may see the available face picture pairs by typing the following command:
lfwpairs_trainsubset = fetch_lfw_pairs(subset='train')
list(lfwpairs_trainsubset.target_names)
3. In conclusion, we will execute the following command to retrieve a list of the same and different individuals:
['Different persons', 'Same person']
Conclusion
In this article, I described common uses of face datasets and introduced four extensive datasets you can use in your face recognition projects:
- CelebA
- Flickr-Faces-HQ Dataset (FFHQ)
- 300W
- AFLW
Finally, I showed how to work with the LFW database to fetch face images and integrate them into your machine learning projects. I hope this will help give you some practical knowledge on how to make use of face datasets in your next projects.
