2019/12/10 by Mkhuseli Ngxande, Ngxande, Mkhuseli, Jules‐Raymond Tapamo +3 · 1 citation
Psychology · #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Sleep and Work-Related Fatigue
paper · pdf · doi:10.48550/arxiv.1912.12123
openalex publication_date 2019/12/10 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
Datasets are crucial when training a deep neural network. When datasets are\nunrepresentative, trained models are prone to bias because they are unable to\ngeneralise to real world settings. This is particularly problematic for models\ntrained in specific cultural contexts, which may not represent a wide range of\nraces, and thus fail to generalise. This is a particular challenge for Driver\ndrowsiness detection, where many publicly available datasets are\nunrepresentative as they cover only certain ethnicity groups. Traditional\naugmentation methods are unable to improve a model's performance when tested on\nother groups with different facial attributes, and it is often challenging to\nbuild new, more representative datasets. In this paper, we introduce a novel\nframework that boosts the performance of detection of drowsiness for different\nethnicity groups. Our framework improves Convolutional Neural Network (CNN)\ntrained for prediction by using Generative Adversarial networks (GAN) for\ntargeted data augmentation based on a population bias visualisation strategy\nthat groups faces with similar facial attributes and highlights where the model\nis failing. A sampling method selects faces where the model is not performing\nwell, which are used to fine-tune the CNN. Experiments show the efficacy of our\napproach in improving driver drowsiness detection for under represented\nethnicity groups. Here, models trained on publicly available datasets are\ncompared with a model trained using the proposed data augmentation strategy.\nAlthough developed in the context of driver drowsiness detection, the proposed\nframework is not limited to the driver drowsiness detection task, but can be\napplied to other applications.\n