NBA Player Facial Recognition and Authentication

• 

NBA Player Facial Recognition and Authentication

Who

atang5, tkim20

Introduction

Facial recognition is an interesting concept that has a lot of practical applications. We are both avid fans of the NBA, so we decided to apply this concept to recognize the faces of NBA players.

The paper we will be implementing is Data-specific Adaptive Threshold for Face Recognition and Authentication, published by researchers at Academia Sinica, Taipei, Taiwan. This paper uses adaptive thresholding to achieve greater accuracy on facial recognition relative to the conventionally used fixed thresholding method and examines its practicality in the context of facial recognition for online registration. User authentication is one of the many practical uses for facial recognition, so we wanted to choose this paper as it illustrates how we could potentially use these technologies to serve a legitimate real-world purpose.

Related Work

The selected paper is considered the most reputable with specific regards to adaptive thresholding in facial recognition. However, there are a select number of other papers which similarly implement adaptive thresholding for the sake of this problem, but none that also implement such using a neural network, nor do any also examine facial recognition in the context of user authentication.

Here is a paper which details the use of adaptive thresholding in person verification. Person verification is slightly different from facial recognition/authentication, in that it attempts to determine whether or not two images are of the same person (regardless of if this image has been seen before). This paper takes its approach from traditional metric learning (ML) approaches in using a locally adaptive decision rule, and is able to achieve greater accuracy on certain benchmark datasets.

This is the official implementation of the research paper on GitHub: link.

Data

We are using a dataset of images of NBA players. We will need to manually preprocess this data to make sure that they are all single face images of the correct player (by label). There are about 520 NBA players in this dataset, with ~70 pictures each (~36,400 total images).

Methodology

The model will implement a convolutional neural network with adaptive thresholding (with fixed thresholding used as a benchmark). The hardest part of this paper, in my opinion, will be understanding the Tensorflow implementation to the extent that we can implement it ourselves in Pytorch.

Metrics

We will measure our success based on the accuracy of our model on our testing dataset. The authors of the paper measured their success by the increased accuracy of adaptive thresholding compared to fixed thresholding, and we will similarly look to both thresholding methods to find an increase.

Base: Implement a convolutional neural network for images (with slightly more functionality than HW2) Target: Base + include adaptive thresholding, with a certain level of accuracy increase below the research paper. Stretch: Target + see how to potentially adjust thresholding for racial biases (touched upon below in ethics section).

Ethics

What broader societal issues are relevant to your chosen problem space?

Facial recognition concerns ethics greatly. Algorithms are known to have racial biases, which would undoubtedly shine bright in identifying people by their faces. The NBA is predominantly black (74.4% to be exact), so we will look to investigate or uncover any inherent biases in these algorithms by training on such a dataset.

What is your dataset? Are there any concerns about how it was collected, or labeled? Is it representative? What kind of underlying historical or societal biases might it contain?

Our dataset of images of NBA players is from Kaggle. According to the user who posted it: "This is an image dataset for all NBA Players from the 2019-20 season. There are about 50 - 80 images for each player and its sorted folder-wise to make labeling easier. I have not filtered out all photos so there may be some photos that may not belong to the player. All these images were scraped from Google images." As such, the biases reflected in the collection of the dataset are directly stemming from those biases in search engines. Already in one of the first datasets alphabetically (Bam Adebayo), I notice there is mistakenly an image from Giannis Antetokumpo. We will make sure to manually clean this dataset, potentially limiting the training dataset to only 30 images per player.

Check-In #2

Reflection linked here.

Final Writeup

Final writeup here

Built With

• python
• pytorch