Detection, Landmarking
Detect faces within a frame, generate key points on the face, and begin tracking dwell time and session time. Pass that detection into our neural networks for our demographics.
Neural Network
Once we detect and landmark, the detection is passed to our neural network to generate a hierarchy of features, to compare to our existing model of what our demographics should look like.
Generate Demographics
From there, we generate age, gender, and emotion for detected person and push these metrics to the dashboard, trigger webhooks, and store in our database.
Accuracy and Metrics
Below are updated statistical analysis of the accuracy of our data points. We improve these daily and will grow the types of metrics we gather as we further develop our technology platform.
| Metrics | Accuracy (%) |
| Happy | 80.5 |
| Angry | 83.3 |
| Surprised | 91.1 |
| Fear | 82.6 |
| Sad | 78.1 |
| Disgust | 98.5 |
| Calm | 80.8 |
| Metrics | Accuracy (%) |
| Male | 91.3 |
| Female | 91.3 |
| Age | +/- 10 years |
| View | +/- 5 degrees (91.3%) |
| Dwell Time | +/- 100 ms |
Uniqueness and Occlusion Algorithms
We use a combination of complex tracking and detection algorithms to maintain a person's session across the screen.
Our complex algorithms can handle occlusions, obstructions, and lost detections, to maintain the uniqueness of a person over the session.
A user's "session" is considered from the time the face is first detected, till an interval of time that the face can no longer be detected.
Low Latency
Our Computer Vision algorithms allow for low latency implementations, on low cost hardware solutions on site, or remotely.
We've optimized our detections to process faces as low as 2 milliseconds per face at a distance of approximately 25-30 feet.
Range
Our Computer Vision technology can be applied to any number of cameras with short and long-range vision.
The quality of our metrics are independent of the range, and purely based on the pixel dimensions of the detections.
Our testing range and minimum detection size is 30x30 pixels, at a distance of 25 feet, and can be extended with higher resolution cameras.
Multi-Platform Support
Our libraries were designed to be easily deployable and support both OpenCV and OpenCL integrations.
On how to get started with our SDKs, please contact us.
