You retain listening to about AI-powered cameras, nevertheless it usually sounds just like the Wizard of Ouncesis working behind the scenes. Now Samsung posted a reasonably detailed rationalization of how the Galaxy S21 collection makes use of AI to create spectacular portrait pictures.
The selfie digital camera processing is fairly simple. First, the software program identifies faces within the picture and marks them for additional processing (that is referred to as segmentation). Second, element in hair, eyes and facial options is enhanced. The digital camera additionally tweaks the white steadiness to attain pure pores and skin tones whatever the ambient gentle.
Portrait mode is rather more concerned. It begins with segmentation so later levels know which components of the picture are people, that are pets and that are simply background. This manner the proper processing will be utilized the place wanted. This segmentation map is used to create a tough “seed map”.
Subsequent is the “tri-map”, which is vital because it highlights the border between the topic and background. Then the matte map traces fantastic element inside that border – it retains hair and facial options from mixing into the background. Lastly, the depth estimation move calculates the gap to things, which might be used to create the shallow depth of discipline impact.
In the event you’ve encountered lower than excellent portrait modes, you’ll know the way usually strands of hair and different fantastic element is blurred into the background, giving the picture a really synthetic really feel. It’s the accuracy of the matte map that units Samsung’s portrait pictures other than the remainder.
Under is an instance tri-map (middle picture), white for the topic, black for the background and gray for unsure areas. On the correct is the ensuing matte map, which refines the topic/background separation.
The cellphone takes the seed maps to use picture enhancements to the topic and blur the background. Then, it provides the matte map to create a pointy division between topic and background. All these layers are processed and mixed into the ultimate picture in three seconds.