...

Please also note that the estimated gaze direction may be a bit unstable (the gaze vectors appearing “shaky”) due to the difficulty of accurately locating the pupils. At the same time, the 3D head pose (head direction) is much more stable. Because people usually turn their head in the direction where they are looking, it may also be interesting to use the head pose as the approximation of the direction of gaze.

...

Here is a code snippet used to roughly calculate screen space gaze by combining data from visage|SDK and external information about screen (in meters) and camera relation which can be used to determine if user is looking at the screen or not (if calculated coordinates are outside the screen range):

   // formula to get screen space gaze

    x = faceData->faceTranslation[2] * tan(faceData->faceRotation[1] + faceData->gazeDirection[0] + rOffsetX) / screenWidth; // rOffsetX angle of camera in relation to screen, ideally 0

    y = faceData->faceTranslation[2] * tan(faceData->faceRotation[0] + faceData->gazeDirection[1] + rOffsetY) / screenHeight; // rOffsetY angle of camera in relation to screen, ideally 0

    // apply head and camera offset

    x += -(faceData->faceTranslation[0] + camOffsetX); // camOffsetX in meters from left edge of the screen

    y += -(faceData->faceTranslation[1] + camOffsetY); // camOffsetY in meters from top edge of the screen

Can you process images of crowds?

24 Jul 2020

visage|SDK can be used to locate faces in group/crowd images and also to perform face recognition (identity) and face analysis (age, gender, emotion estimation). Such use requires particular care related to performance issues as there may be many faces to process. Some initial guidelines:

• Face tracking is limited to 20 faces (for performance reasons). To locate more faces in the image, use face detection (class FacialFeaturesDetector).

• There is a performance trade-off in face detection: detecting smaller faces in the image is computationally more expensive.

• For the face to be detected, it should be about 30 pixels wide in the image. For face recognition and analysis (age, gender, emotion), higher resolution is better, ideally 100-150 pixels wide.

Troubleshooting

I am using visage|SDK FaceRecognition gallery with a large number of descriptors (100.000+) and it takes 2 minutes to load the gallery. What can I do about it?

...