2
do better? Can we automate this problem in a bid to reduce human dependency and to
simultaneously obtain better results?
One must acknowledge that aging of face is not only determined by genetic factors
but it is also influenced by lifestyle, expression, and environment [1]. Different people
of similar age can look very different due to these reasons. That is why predicting age
is such a challenging task inherently. The non-linear relationship between facial images
and age/gender coupled with the huge paucity of large and balanced datasets with cor-
rect labels further contribute to this problem. Very few such datasets exist, majority
datasets available for the task are highly imbalanced with a huge chunk of people lying
in the age group of 20 to 75 [3]-[5] or are biased towards one of the genders. Use of
such biased datasets is not prudent as it would create a distribution mismatch when
deployed for testing on real-time images, thereby giving poor results.
This field of study has a huge amount of underlying potential. There has been an
ever-growing interest in automatic age and gender prediction because of the huge po-
tential it has in various fields of computer science such as HCI (Human Computer In-
teraction). Some of the potential applications include forensics, law enforcement [1],
and security control [1]. Another very practical application involves incorporating these
models into IoT. For example, a restaurant can change its theme by estimating the av-
erage age or gender of people that have entered so far.
The remaining part of the paper is organized as follows. Section 2 talks about the
background and work done before in this field and how it inspired us to work. Section
3 contains the exact technical details of the project and is further divided into three
subsections. Section 4 talks about the evaluation metric used. Section 5 presents the
various experiments we performed along with the results we obtained, and finally sec-
tion 6 wraps up the paper with conclusion and future work.
