Results
- Probability of right detection: up to 0.99 (99.99%)
- Probability of right classification: up to 0.96 (96%)
- Autonomous mobile face detection and recognition
- Our own algorithm for face geometrical normalization
- Optimized to reduce the consumption of energy as much as possible
Key Technologies involved
- C++
- JavaScript / ECMAScript
- Python
- OpenCV / OpenCV ML / OpenCV Light
- Face Lib
- Android / Windows / Linux
- Dlib
- QT / QML
- Java / JNI
- Git
Introduction
The guy in the image below is Kostia, one of our talented developers. As you can imagine, the strange shapes drawn on his face are not a clumsy post-impressionist makeup attempt but the bounding boxes of our facial recognition software.
That’s right, we use our own programmers as guinea pigs to test our products.
Brutal..
Our goals
The idea was to develop a face detection and recognition tool directly in mobile phones, without the need to establish a connection to the server or to use the internet.
The second goal was to create an app that didn’t drain too much energy from the cell phone battery.
The Apro performance
The application has shown excellent results for both recognition and classification tasks:
Probability of right detection: up to 0.99 (99.99%)
Probability of right classification: up to 0.96 (96%)
To reach these results, our R&D project included not just coding but a significant amount of interactions with real people inside a controlled environment.
Who says that a developer’s life is always lonely?
The app features
Our solution has its own image normalization algorithm that provides better quality in case of not straight face orientation.
It is also smartly optimized to reduce the consumption of energy as much as possible.
A face recognition model is stored, trained, and updated directly in a smartphone. This means that all the set of operations proceeds autonomously and independently from external services.