Noah Kalina has been taking daily photos of his face for the past 21 years and periodically makes timelapse compilations of the photos. For his latest video he collaborated with Michael Notter on a video called 7777 Days where they used machine learning to align and upscale the last 7777 photos into a two minute video.
In a first step, Michael used the machine learning library dlib (http://dlib.net/) and some custom Python code to detected in each of Noah’s photos 5 face landmarks (i.e. both eyes, the nose and the two corners of the mouth). These landmarks were then used to align the faces in all photos, so that the eyes and corner of the mouth were horizontally oriented and always an equal distance apart. After that, some small image intensity correction were applied to make very dark images a bit brighter and very bright ones a bit darker. This was followed by an upscaling of all images (where needed) to a 4K resolution.
In a second step, once the faces were upscaled and aligned, Michael looped through all of the images and averaged them with a sliding window approach: Each frame in the video shows the average face of the last 60 faces. Or in other words, each frame shows the ‘average Noah’ over the last 2 months. With a video frame rate at 60Hz, this means Noah ages in this video 2 month every second, or 10 years every minute.
You’ve probably already seen this guy’s face in videos like this, but this is the most complete and definitive version yet. My only complaint would be whatever technique they used to blur the backgrounds. It makes the whole video look too artificial and like his face was also machine generated. You also lose the detail in the hair and all I really care about is seeing that sweet ‘do change.
Keep going for the full video.