Frustrations from the current state-of-the-art in automatic face recognition.
For the past few weeks, I have been carrying out some forensic face recognition experiments based on the CMU MultiPIE dataset. The objective of these experiments is to test a new framework for forensic face recognition. Therefore, in these experiments, I use a commercial face recognition system as a “black box”. In other words, in these experiments, I am not concerned with how this commercial face recognition system works but I aim to present a framework for applying automatic face recognition systems to forensic cases.
Today, I finished the last part of the experiment and I feel so frustrated from the results that I have decided to express my frustrations in this blog post. To be honest, I had never been so frustrated in the past few years.
I have the following appeal to all the researchers (including myself) involved in face recognition:
- Try to venture outside the comfort zone of controlled face image dataset. Most face recognition algorithm developers feel safe to test their algorithms using only standard face image dataset captured in controlled (pose, illumination, etc) environment. Please have the courage to test algorithms on uncontrolled images and face the reality.
- Stop pursuing research to improve face recognition accuracy from 96% to 98% on a controlled face image dataset. We have already lost the past decade doing such incremental research.
- Stop using idiosyncratically chosen facial features which can hardly be attributed to an individual’s identity. Such rouge facial features are often found wandering in the feature space when subject to facial pose and illumination variation.
It is not difficult to understand why the professionals involved in forensic face recognition cases prefer to use trained human experts for face matching. Today’s frustration has a bright side as well. I now feel happy to be exploring a dilapidated research avenue. Not only more reasons to be frustrated but also more reasons to try out new ideas and challenge the conventional wisdom.