Background

Emotions are hardwired into our brains at birth and manifest as facial expressions. Research has long proven their universality, regardless of age, gender or race. Facial expressions are classified as joy, surprise, sadness, disgust, fear and anger. Findings on contempt are ambiguous and less clear. To maintain our highest standards of accuracy, we do not include contempt in our algorithms.

Technology & Intellectual Property

We dedicated years in developing our emotion recognition software. We built our technology from the ground-up and we don’t reply on licensing third party components. We invested heavily into collecting our own proprietary datasets and we own 100% of our Intellectual Property. We created novel techniques that use Deep Learning architectures, which employ Convolutional Neural Networks (CNN) to train our algorithm and deploy it to the real world for continuous learning.

Deep Learning

Deep Learning

Our holistic facial expression recognition methodology imitates human vision and allows our algorithm to learn prototypic expressions directly from the face instead of relying on decomposed action units.

In addition, our methodology computes both shape and texture information which results in unparalleled accuracy under variant uncontrolled environments. Our Deep Learning approach allows many benefits, mainly around performance, over any other methods.

Learning Through Conv Nets

We employ cutting edge and proprietary Deep Learning-based algorithms to achieve our superior classification accuracy across the supported metrics via Convolutional Neural Network (CNNs). Challenging covariate factors such as pose and lighting variations are mitigated by employing robust face tracking, pose normalization and synthesis, and local lighting invariant feature descriptors. Misclassification is minimized through a rigorous classification scheme comprised of decision rule/filtering, classification, and verification stages. Our classifiers are trained on a large and diverse corpus of images to account for variations in appearance across different settings.

Dataset Creation

We collected our proprietary dataset and trained our algoritms using Deep Learning on the most versatile physical and environment variables that are critical to achieving the highest level of emotion recognition accuracy.

These variables include dataset collection from major races and ethnicities, among different age groups and genders, under posed environments and in-the-wild, using different lighting conditions, in both frontal and non-frontal head poses, and with the use of different attributes (glasses, hats…etc).

In addition, we created an Deep Learning “emotional vocabulary” that allows easy input and training of additional case-specific datasets for improved accuracy.

Benefits of the Deep Learning approach

Accuracy

Accuracy

EmoVu was designed with Embedded Systems in mind. The ultra-lightweight software runs locally, requires minimal processing power and is optimized for integration into most embedded platforms (MCU, SoC, DSP…).

Speed

Speed

We created novel methodologies to train and deploy our technology algorithms using Convolutional Neural Networks (CNN’s) as a Deep Learning architecture to achieve the most accurate facial expression recognition software available.

Customization

Customization

Our underlying technology revolves around a structured framework that we initially conceived as an Emotional Vocabulary. This allows fast and easy optimizations as needed based on race, environment and context.

Continuous Learning

Continuous Learning

Our “live” algorithm learns from every image or video it processes. This ensures continuous learning, which improves performance and accuracy on an everyday basis