Face Recognition recommendations

Recommendations on characteristics for choosing the cameras:

• Minimal required resolution is 2 MP. Since higher resolution requires more network, HDD and CPU/GPU resources - it is not recommended to use resolution higher resolution where it is not required.
  Required minimal PPM value in recognition zone is 250. For stable work it's better to consider up to 500 PPM. Detector quality decreases the PPM required for the recognition, but increases the load on the server.
• Matrix size from 1/2.8" (minimal required) to 1/1.8" depending on the lighting conditions.
• Aperture from F1.4 (minimal required) to F1.2 depending on the lighting conditions.
• Possibility of shutter speed control and to make it fixed. Moving objects have to be clearly visible.
• Varifocal lens at least x4 optical zoom. The distance between the pupils in the image must be at least 60 px. The size of the face in the frame has to be not less than than 128 px.
• The more light sensitive will be the matrix - the better. FTC cameras are good option.
• RWDR 120 dB (camera hardware WDR), better 140 dB.
• Onvif-compliance (at least S).

Recommendations on camera settings:

• Camera bitrate has to be sufficient for required resolution, FPS an motion intensity. The camera picture has to be stable, without noise and glitches.
• Codec has to be H.264.
• The minimum stream compression should be selected.
• Disable autofocus on camera.
• Disable OSD text on camera if possible.
• Disable camera software picture enhancements. Consider using only RWDR if WDR is required.

• The depth of field should, if possible, cover the entire area of the intended detection. Sharpness should be tuned to get the highest quality picture in the area of the best conditions for face recognition.

• The less GOP - the better.
• Stream compression should be set on minimum.

Other factors affecting the Face Recognition performance:

• The lighting in the scene of recognition should be sufficient and equal. At the night time, there should be no overexposure of the objects and at daytime there should be no shadows on faces. Faces in the image should be sufficiently contrasting, illuminated and clearly visible to the naked eye.
• The positioning of the camera. The shooting direction should be in such a way that people's faces look directly into the camera lens. Ideal vertical and horizontal angles are up tp 15 degrees, permissible value is up to 30~35 degrees.
  

• For high-quality operation of the module, the camera should be installed in places with a centralized and structured passage area, for example, the front door to the building. Wide halls with chaotic movement of people are a bad place to install a camera.

• It is recommended to install multiple cameras to capture a wide area.
• Use professional designing tools for camera installations, such as JSVG - The use of a lens calculator 

Recommendations on photos of persons to be added in the face database:

• It is possible to upload several photos for one person, one of which must be taken in full-face, and the rest are allowed to rotate no more than 30 degrees vertically or horizontally.
  Photos in which the person is depicted half-face, will significantly reduce the probability of this person recognition.

• If a person wears glasses, then upload an additional photo with glasses to improve the recognition.
• Photos on which a person's face is blurry, lighted or in shadow will significantly reduce the probability of recognition.
• The face on the photos should not be smaller than 128x128 px.
  
  
  See examples there

Most common issues

• Issue - Overexposured entrance, strong backlights
  Solution - Hardware WDR (RWDR)
  
  
  
• Issue - Camera installation on a long bracket from the ceiling might lead to fluctuations, for example, from air currents in the room.
  Solution - applying additional measures for fastening - for example, additional supports / stretches.
• Issue - Module is capturing blurry faces.
  Solutions - the shutter speed increase.
• Issue - Module does not capture faces at all.
  Possible solutions:
  • increase the maximum bitrate or decrease the resolution / FPS (if that high values are not required)
  • double-check the detector set up - https://www.dssl.ru/files/trassir/manual/en/setup-face-recognizer-description.html
  • double-check the installation angles - probably they are too much
  • double-check the lighting conditions in the scene - faces might not be lighted enough or opposite to that - overexposured
  • maybe more cameras are required to cover a wide area
  • maybe the cameras characteristics or settings are not sufficient

Examples of good angles and image quality: