In retail environments, it is important to acquire information about customers entering in a selling area, by counting them, but also by understanding stable traits (such as gender, age, or ethnicity) and temporary feelings (such as the emotion). Anyway, in the last year, due to the COVID-19 pandemic, it is becoming mandatory to wear a mask, covering at least half of the face, thus making the above mentioned face analysis tasks definitely more challenging. In this paper, we evaluate the drop in the performance of these analytics when the face is partially covered by a mask, in order to evaluate how existing face analysis applications can perform with occluded faces. According to our knowledge, this is the first time a similar analysis has been performed. Furthermore, we also propose two new datasets, designed as extensions with masked faces of the widely adopted VGG-Face and RAF-DB datasets, that we make publicly available for benchmarking purposes. The analysis we conducted demonstrates that, except for gender and ethnicity recognition whose accuracy drop is quite limited (less than 10%), further investigations are necessary for increasing the performance of methods for age estimation (MAE drop between 4 and 10 years) and emotion recognition (accuracy decrease between 45% and 55%).
Performance Assessment of Face Analysis Algorithms with Occluded Faces
Greco, Antonio;Saggese, Alessia;Vento, Mario;Vigilante, Vincenzo
2021-01-01
Abstract
In retail environments, it is important to acquire information about customers entering in a selling area, by counting them, but also by understanding stable traits (such as gender, age, or ethnicity) and temporary feelings (such as the emotion). Anyway, in the last year, due to the COVID-19 pandemic, it is becoming mandatory to wear a mask, covering at least half of the face, thus making the above mentioned face analysis tasks definitely more challenging. In this paper, we evaluate the drop in the performance of these analytics when the face is partially covered by a mask, in order to evaluate how existing face analysis applications can perform with occluded faces. According to our knowledge, this is the first time a similar analysis has been performed. Furthermore, we also propose two new datasets, designed as extensions with masked faces of the widely adopted VGG-Face and RAF-DB datasets, that we make publicly available for benchmarking purposes. The analysis we conducted demonstrates that, except for gender and ethnicity recognition whose accuracy drop is quite limited (less than 10%), further investigations are necessary for increasing the performance of methods for age estimation (MAE drop between 4 and 10 years) and emotion recognition (accuracy decrease between 45% and 55%).I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
