mEBAL2 [1] is a new multimodal database for eyeblink detection and attention level estimation obtained from an e-learning environment, a new upgraded version from the well-known database mEBAL [2]. This database is the largest existing public eyeblink database, with 21,100 labeled image sequences (10,550 eyeblinks and 10,550 no-blink events) and students’ cognitive activity labels from 180 different students, while conducting a number of e-learning tasks of varying difficulty or taking a real course on HTML initiation through the edX MOOC platform. This information is avalible on this web [Download Database].
mEBAL2 has been published in the journal Pattern Recognition Letters in June 2024. A preprint is also available on arXiv.
mEBAL2 has improved previous databases in terms of sample numbers in eye blink and cognitive information. In particular, three different sensors are simultaneously considered: Near Infrared (NIR) and RGB cameras to capture face gestures, also an Electroencephalography (EEG) band to capture the cognitive activity of the user and eyeblink events.
Also, this new version of mEBAL comes with significant improvements:
- Additional 120 users with 7.550 eye blink events, making it three times bigger, and more than 2,5 million frames recorded from 180 students. Furthermore, a new real e-learning environment has been added for monitoring purposes, specifically a MOOC.
The following sections describe the motivation, tasks, sensors, public database, and the mEBAL2 database will be shared with the community to move forward in this area:
Eyeblink has proven to be a valuable indicator in various fields such as ocular activity, attention, fatigue, emotions, etc., for this reason, eyeblink detection based on image processing has become essential regarding applications involving human behavior analysis.
Eyeblink detection is a useful tool to improve e-learning platforms [3, 4] to get high-quality online education, for at least two reasons. First, since the 70s, there are studies relating the eyeblink rate with cognitive activity like attention. Recent research [5, 6] suggest that lower eyeblink rates can be associated with high attention periods, while higher eyeblink rates are related to low attention levels. And second, blink detection can be used in the detection of fraud/cheating/lies and combined with other features like heart rate, gaze tracking, micro-gestures or blood oxygen saturation, can improve the security of e-learning platforms.
However, the state of the art demonstrates that public eyeblink detectors based on image processing are far from resolving the eyeblink detection problem. At the moment, there are very few public data-driven algorithms (e.g., neural networks), mainly because of the lack of large eyeblink databases. Most existing datasets useful for research in this area, have only a few hundred samples, representing a strong restriction to train data-driven approaches (e.g. deep learning). Also, current public databases restrict their samples to RGB cameras, without using other sensors that have proven to be useful in similar tasks such as NIR cameras in gaze tracking or iris and pupil detection. In the mEBAL2 [1] article it is demonstrated that the approaches trained with both spectra (visible and NIR) have a good generalization capacity for unseen scenarios, showing the how useful the NIR camaras can be.
For this reason, we hope that mEBAL2 will be valuable for the scientific community, thanks to the multimodal nature: Cognitive Activity and EyeBlink Detection.
The database was divided into two groups. The first group of 60 students did a series of tasks that were carefully designed to reach certain goals, and the second group of 120 students did a real lesson from a MOOC, entitled “Introduction to Development of Web Applications” (WebApp), which is available in the edX platform.
The tasks for both groups were designed with two goals. First, to generate changes in the students’ cognitive activity such as mental load, attention, visual attention, etc., looking to cause variations of the eyeblink rate. And second, to generate a realistic setting of online assessments. The tasks can be categorized into five groups:
-
Enrollment form: Student’s data are obtained here. This is a simple task that targets a relaxed state with attention levels between normal and low.
-
Logical questions: These require more complex interactions, and some of them include crosswords and mathematical problems, for the first group. For the MOOC course, some of the activities involve writing HTML codes and generating more efficient ones.
-
Visual tasks: These demand visual attention from the students under different situations, such as: watching prerecorded classes, describing images, detecting errors in HTML code, etc.
-
Reading tasks: Reading documents has proven to have an impact on eyeblink rates and it’s highly common in e-learning environments.
-
Multiple choice questions: These are essential to help evaluate the students on assessment platforms and most Learning Management Systems provide templates to perform these assessments.
mEBAL2 contains synchronized information from multiple sensors while the students use an interface designed for e-learning tasks. Therefore, a multimodal acquisition framework was designed to monitor cognitive and eyeblink activity based on the edBB platform [3, 4], an e-learning platform for remote education assessment:
 |
|---|
| Setup used on mEBAL2 and the eyeblink diagram of the acquisition from mEBAL2 |
The acquisition setup uses the following sensors:
-
An Intel RealSense (model D435i), which contains 1 RGB and 2 NIR cameras. The NIR cameras are monochrome and sensitive in the visible spectrum and NIR, following the sensitivity curve of the CMOS sensors. The 3 cameras are configured to 30 Hz (one frame every 33ms) and 1280 × 720 resolution. It's known that an average blink takes 100ms to 400ms, therefore, an eyeblink can take between 3 to 13 frames.
-
An EEG headset by NeuroSky, which measures the power spectrum density of 5 electroencephalographic channels (α, β, γ, δ, θ). EEG measures the voltage signals produced usually by synaptic excitations of the dendrites of pyramidal cells in the top layer of the brain cortex. Eyeblinks introduce artifacts that can be easily recognized in EEG signals. In this dataset, the EEG band was used to generate the initial ground truth data necessary to label the eyeblink events. We have made a manual refinement of these eyeblink candidates detected by the band to eliminate false positives. These refined eyeblinks will be used as eyeblink groundtruth. The sampling rate of the band is 1 Hz.
The following table shows the sensors and the information captured:
 |
|---|
| Sensors included in the mEBAL2 framework |
mEBAL2 includes 21,100 events (10,500 blinks and 10,500 no-blinks) from 180 students/sessions. The session duration varies from 15 to 40 minutes. Each eyeblink event has 19 frames using three cameras: one RGB, and two NIR cameras. This database contains 2,405,400 frames (3 cameras × 19 frames × 21,100 events × 2 eyes), making it the largest existing eyeblink database.
Therefore, mEBAL2 provides a dataset consisting of 540 long-duration videos (1 RGB video and 2 NIR videos per session). Each video comes along with the facial bounding box information, 68 facial landmarks, and cropped eye regions for each frame. Furthermore, the dataset includes timestamps for both eyeblink and no-eyeblink events and a total of 21,100 cropped samples. Additionally, the dataset provides EEG band information, including attention level, meditation level, 5 electroencephalographic channels, and eyeblink intensity measures.
Additionally, the new mEBAL2 contains variations on illuminations, poses, distances between user and camera, objects over the face (glasses, hair, hand occlusion, etc.), physical activity, and other naturally-occurring factors. The next figure shows some examples from mEBAL2:
 |
|---|
| Different examples from mEBAL2. (top) Sequence images with variations in illumination, posing, and distance to the camera. (bottom) Examples of eyeblink and no-blink with RGB and NIR images. |
The eyeblinks were labeled using a semisupervised approach. For that labelling we first used as candidates for ground truth the eyeblink information provided automatically by the EEG band, and then a human checked manually all the detected events.
The next table shows the most popular eyeblink detection databases:
 |
|---|
| Existing databases for Eye Blink detection. |
mEBAL2 is the largest eyeblink public database.
-
Download license agreement, send by email one signed and scanned copy to atvs@uam.es according to the instructions given in point 2.
-
Send an email to atvs@uam.es, as follows:
Subject: [DATABASE: mEBAL2]
Body: Your name, e-mail, telephone number, organization, postal mail, purpose for which you will use the database, time and date at which you sent the email with the signed license agreement.
-
Once the email copy of the license agreement has been received at ATVS, you will receive an email with a username, a password, and a time slot to download the database.
-
Download the database, for which you will need to provide the authentication information given in step 4. After you finish the download, please notify by email to atvs@uam.es that you have successfully completed the transaction.
-
For more information, please contact: atvs@uam.es
-
[1] Daza, R.; Morales, A.; Fierrez, J.; Tolosana, R.; and Vera-Rodriguez, R. 2024. mEBAL2 Database and Benchmark: Image-based Multispectral Eyeblink Detection. In Pattern Recognition Letters. [pdf]
-
[2] Daza, R.; Morales, A.; Fierrez, J.; and Tolosana, R. 2020. mEBAL: A Multimodal Database for Eye Blink Detection and Attention Level Estimation. In ACM International Conference on Multimodal Interaction. [pdf]
-
[3] Hernandez-Ortega, J.; Daza, R.; Morales, A.; Fierrez, J.; and Ortega Garcia, J. 2019. edBB: Biometrics and Behavior for Assessing Remote Education. In AAAI Workshop on Artificial Intelligence for Education. [pdf]
-
[4] Daza, R.; Morales, A.; Tolosana, R.; Gomez, L. F.; Fierrez, J.; and Ortega-Garcia, J. 2023. edBB-Demo: Biometrics and Behavior Analysis for Online Educational Platforms. In Proc. AAAI Conf. on Artificial Intelligence (Demonstration). [pdf]
-
[5] Daza, R.; DeAlcala, D.; Morales, A.; Tolosana, R.; Cobos, R.; and Fierrez, J. 2022. ALEBk: Feasibility study of attention level estimation via blink detection applied to e-learning. In Proc. AAAI Workshop on Artificial Intelligence for Education. [pdf]
-
[6] Daza, R.; Gomez, L. F.; Morales, A.; Fierrez, J.; Tolosana, R.; Cobos, R.; and Ortega-Garcia, J. 2023. MATT: Multimodal Attention Level Estimation for e-learning Platforms. In: Proc. AAAI Workshop on Artificial Intelligence for Education. [pdf]
For more information contact Aythami Morales, associate professor UAM at aythami.morales@uam.es