From Pavements to Patterns: Unravelling Trends in Age, Gender, and Attire for Societal Awareness

Abstract

The recent developments in computer vision and deep learning have opened a range of different application domains, such as object detection, facial recognition, medical imaging, and smart cities. The focus of this thesis is on object detection. Specifically, we investigate apparent age estimation and gender classification with apparel detection and classification. The existing literature does not address the concurrent estimation of apparent age, gender classification, and attire detection in the wild. To address this limitation, we create a novel dataset comprising age, gender, and attire labels. After labelling the dataset, we apply face alignment and image enhancement techniques to improve the existing age estimation and gender classification algorithms' efficacy. Our experimental results show that RetinaNet has the highest precision (1.0) and recall (0.97) among the analysed algorithms for face detection on our test set. Furthermore, we observe that varying the area around the face during the training phase impacts the efficacy of the analysed algorithms. We illustrate that incorporating ~40% of the area around the face coupled with image enhancement via GFP-GAN maximises the efficacy of the analysed algorithms for apparent age estimation and gender classification. We show that FRCNN is the best performing algorithm for estimating ages (mean absolute error of 6.85) on our test set, whereas, DETR-Deform is the best performing algorithm for gender classification (average precision of 0.89 for female and 0.91 for male) and attire detection (mean average precision of 0.83). The experimental results demonstrate the practicality of our approach in real-world settings.