The features “adapt my face model” does not work…
UPDATE : a new model management has been introduced since the 2.5 version of KeyLemon in order to correct the problem described in this article. The new model management creates multiple models for each lighting environment.
In this Blog Post we would like to explain you how the “adapt my face model” features works or how to use it correctly.
The “adapt my face model” feature permits to be recognized under different environmental conditions. It becomes very useful when you are using KeyLemon in different lighting conditions. Let’s imagine you are traveling in a train and you would like to use your laptop. KeyLemon might have some difficulties to recognize you as lighting conditions will change frequently. In this case, adapting your face model will make sense.
However, in some cases, the adapt model does not work as expected. This happens when the face model is adapted in bad conditions, like strong or weak lighting. Let’s have a look at the illustration below to have a better understanding how the “adapt my face model” works.
A model is a mathematical calculation of points in a cloud. When you adapt the model, a mix of both will be generated. The corresponding new cloud of points does not reflect a “real” image but it is a sort of mathematical mean of the two clouds. Each model gets a bit degraded. However, the new model should be the best compromise to be recognized in the new and old situation. If you redo an “adapt my face model”, the already adapted model will be used as the base model.
If the original model and the adapted one are too “far away”, the mix will be an incoherent model. As a consequence, the new adapted model won’t work and you might be disappointed. This happen if a face model was created in quite well conditions (e.g. at the office with uniform lighting conditions) and the adapted in a poor environment with directional lighting.
Adapting the model quite a lot could result degradation. The calculation of the cloud of points will be less and less linked to a real image. If several adapts have been done and the results are not satisfying, we recommend to delete the model and start with a new one.
Stay tuned, new information will follow…