Skip to content

Accuracy & Guidelines

TL;DR: VitalLens achieves high fidelity on high-quality video inputs, but input quality matters.

In our study conducted on a diverse benchmark of 422 unique individuals, we observed a mean absolute error of:

  • Heart Rate: 1.57 bpm
  • Respiratory Rate: 1.08 breaths/min
  • HRV-SDNN: 10.18 ms

Guidelines for Best Accuracy

rPPG estimates vitals from small colour changes in the skin. Anything else that changes the colour or brightness of the face during the measurement may be read as part of the signal. To maximize the confidence scores and accuracy of the estimates returned by the API, follow these guidelines.

The four factors with the largest effect are:

  1. Enough light on the face.
  2. Steady light: No flicker and no changing colours.
  3. Short distance: Avoid distances of more than 0.5 m between the face and the camera.
  4. Little movement of both the subject and the camera.

1. Enough Light

  • Brightness: The face needs to be well lit. In dim light the camera raises its sensitivity, which adds image noise, and video encoders compress dark, noisy images more heavily. Both reduce the quality of the signal.
  • Camera capability: How much light is enough depends on the camera. A camera with a larger sensor or better low-light performance needs less light than a basic webcam. Visible grain in the image indicates that more light is needed.
  • Direction: Light the face from the front or the side. Avoid strong backlight such as a window behind the subject, because the camera then exposes for the bright background and underexposes the face.

2. Steady Light

The face should be lit mainly by one light that does not change in brightness or colour during the measurement.

  • Flicker: Older fluorescent tubes and some LED lights flicker at frequencies that interfere with pulse detection. Use a different light source if the lighting flickers.
  • Screens: Avoid sitting in a dark room that is lit only by a computer monitor or television. Screens refresh at specific rates and emit changing colours that corrupt the rPPG signal on the face.
  • Moving light and shadows: Avoid light that changes during the measurement, for example from passing clouds, moving shadows, or a flashing display.
  • Even illumination: Ensure the face is lit evenly. When one side of the face is lit more strongly than the other, small movements cause large brightness changes. Shadows moving across the face create artifacts.
  • Camera auto-adjustment: Cameras adjust exposure and white balance automatically, and these adjustments change the brightness and colour of the whole image. If your application can lock exposure and white balance, lock them once the face is in position. Otherwise, wait for the image to settle before starting.

3. Distance and Framing

  • Distance: Avoid distances of more than 0.5 m between the face and the camera. A closer face covers more pixels, which gives a stronger signal. The whole face must remain in the frame.
  • Orientation: The subject should face the camera directly.
  • Framing: Include the upper chest in the frame.
    • Face Only: Sufficient for Heart Rate.
    • Face + Upper Chest: Required for Respiratory Rate to detect breathing motion.

4. Motion

While VitalLens is robust to natural movement, signal quality is directly related to stability. The more motion, the lower the confidence score and accuracy of the estimates.

  • Camera Stability: Handheld use is supported, but using a tripod or stable surface will always yield the highest accuracy.
  • Subject Movement: Normal behavior (like blinking or shifting weight) is fine. However, significant activity impacts the signal:
    • Talking/Chewing: Adds noise that may slightly degrade precision.
    • Walking/Exercise: Heavy body motion can overpower the rPPG signal and should be minimized where possible.

5. Video Quality

  • Compression: Avoid heavy video compression (low bitrate). rPPG relies on imperceptible skin color changes. Compression algorithms are designed to remove exactly these "redundant" color details.
  • Resolution: While the API downsamples images internally, inputting uncompressed or high-bitrate video (even at 480p) is preferred over 4K video with high compression artifacts.

Checklist

  • The face is well lit, from the front or the side, without strong backlight.
  • The lighting does not flicker or change colour, and no screen is the only light source.
  • The distance between the face and the camera is not more than 0.5 m, and the whole face is in the frame.
  • The subject and the camera are still.
  • The video is not heavily compressed.

For a deeper technical analysis of how environmental factors impact signal integrity, read our technical paper.