Extra Review & Study on Paper Assignment for Deep Learning class in SKKU -
"Face Quality Estimation and its Correlation to Demographic and Non-Demographic Bias in Face Recognition (Philipp Terhorst, Jan Niklas Kolf, Naser Damer, Florian Kirchbuchner, Arjan Kujiper)"

본문 내용에 대해 [Page - Paragraph - Line] 표기를 하고 있습니다.
아니 근데 벨로그는 한국인만 쓰나요?

---

1. Introduction

[1-2] Face recognition systems are spreading worldwide and have a growing effect on everybody’s daily life. Furthermore, these systems are increasingly involved in critical decision-making processes, such as in forensics and law enforcement. Current biometric solutions are mainly optimized for maximum overall accuracy [31] and are heavily biased for certain demographic groups [38, 4, 16, 40, 6, 18].

This is the 'big' problem that is being faced in the real world.

Current biometric technology focues on maximum overall accuracy as a measure, and in the process, overlooks biases on certain groups - in this case, demographic groups.

For general idea - how can biometric technology have demographic biases?

Global population is not distributed evenly to all existing ethnicities - for example, in US, 60.1% of its population are classified as White, 18.5% are classified as Hispanic, 12.2% are classified as Black, 5.6% are classified as Asian.

[Image Source : Visualizing America's Population By Race]

But you know Deep Learning and Neural Networks - the Networks train based on the training data. If we were to train a model with training data that depicts this unbalanced racial populations (ex, 60 images of White, 18 images of Hispanic ... ), the model would be biased towards specific characteristic features of whites on facial recognition - which may not be common in other races. I mean, I look nothing like a Caucausian - and my feeling would be a little hurt if I were to be classified as non-human, or if my face were not recognized because I don't have 'most common' features of training data.

Yet, the model's accuracy would be high, as it focuses on maximum overall accuracy, than being able to recognize every person of every ethnicity and different characteristics.

[1-2-8] The performance of face recognition is driven by the quality of its captures [5]. Biometric sample quality is defined as the utility of a sample for the purpose of recognition [24, 39, 17, 5] and is crucial for many applications. Recent work [47] has shown that the accuracy and the robustness of face quality estimation can be enhanced drastically by adapting the face quality assessment algorithm to the deployed face recognition model. However, this can lead to biased face quality assessment algorithms as well.

• Biometric sample qaulity : utility of a sample for the purpose of recognition

In facial recognition, this would concern the quality of a facial image for training. Images, or especially in videos, are not always the best for facial recognition.

Also for general idea, I use Face ID for unlocking my phone.
Although I do not know what features or characteristic of my face my Face ID focuses on to evaluate a face, and unlock itself, Face ID tends to work best when I hold it as parallel to my face as I can, without anything near my facial area.

It can't identify me when I wear glasses, or when I look into the camera in a weird angle, or when I am moving and so on.

So this seems natural, right? Because the camera can't really see my bare face at its best angle (which it was trained with) although it IS ME.

Similar applies to Face Recognition.

Some facial images are just not optimal for training - and this "quality" of images are referred to as Face Quality

Image Source : Deep Face Quality Assessment (Agarwhal, V.)

So Face Quality, or Face Image Quality is a subject for assessment too, as seen in this image. Images with

Goal of Paper : pointing out that current face image quality assessment approaches have to deal with similar bias-related problems than in face recognition - the quality of a face image point out a biased ground of a faulty decision

2. Related Work

2. 1. Bias in face recognition

2. 2. Face quality assessment

3. Evaluated face quality assessment solutions

3.1. COTS

3.2. Best - Rowden

3.3. FaceQnet

3.4. SER-FIQ

4. Experimental Setup

Database

Evaluation metrics

Face recognition networks

Investigations

5. Results

5.1. Face quality assessment performance

5.2. Identifying biases in pose, ethniticy, and age

5.3. The correlation study - bias versus quality

COTS

Best - Rowden

FaceQnet

SER-FIQ

Summary

6. Conclusion