The Face Change Dataset consists of frontal images of a face captured several times a day for several months. The goal was to investigate if any systematic shape or color changes could be detected when analyzed over the course of a day, week, or month.
Inspired by: Eulerian Video Magnification for Revealing Subtle Changes in the World we investigated an eigenface decomposition of the variations in appearance.
To apply the method, it was necessary to register the faces. This was done using a facial feature detector, estimating the affine transformation that mapped the detected features onto a reference face, and mapping all of the pixels into a standard position. A mask was then applied to remove hair and background. The eigenface method was used to extract eigenmodes related to color and shading. The weights of the eigenmodes were then investigated as a function of times of day (4 times), week, and month (28 days).
This approach was not as impressive as the Eulerian Video Magnification approach.
We found a slight effect for a few eigenmodes when viewed with a large weight
magnification. Here, you can see 2 examples that we have characterized as Happy:Sad
We also found one case of daily variation (images from roughly 8:00, 13:00, 19:00, and 24:00),
with the variation from early morning stressed through to awake, and then tired and sleepy:
However, in general, no temporal variation was found, with a large variation in weights associated with each hour/day of week/day of month. We did not see any color effects. Maybe future researchers will find some with an alternative analysis.
The dataset consists of several hundred photos of three volunteers over several months. We aimed to acquire images at 4 times a day, but there are some variations in times, and missed times. Volunteer p2 only recorded at one time each day. The dataset is available for other researchers below, and is based on photos volunteered by Sun Yifei, Xu Rongxian, and Bob Fisher. Ethical approval was obtained for the capture of these images.
Any publications that use this data should cite the attached PDF: Y. Sun; "How Does Your Face Change Over Time?", MSc dissertation, School of Informatics, University of Edinburgh, 2021.
P1's images are named 2021-CC-DD HHMMSS.jpg. The images are in 4 subfolders (10,14,20,24) according to the approximate time of day when the images were captured). P2's images are named 2021-CC-DD HHMMSS_Convert.jpg (images with HHMMSS=190000 had an estimated time). P3's images are named PersonX_DDCC21_HHMM.jpg. Here C is for month, D for day, H for hour, M for minute, and S for second.
Email: Prof. Robert Fisher at rbf -a-t- inf.ed.ac.uk.School of Informatics, Univ. of Edinburgh