Single Image Based Infant Body Height and Weight Estimation

Abstract

The collection of infant body data such as height and weight is a useful mean of tracking its growth and wellness. The contact-based measurements using height and weight scales are manual and cumbersome, camera-based methods were proposed to obtain features from face or body for height and weight estimation. In this paper, we created a clinical dataset including 200 newborns collected at obstetrics, and benchmarked four convolutions neural networks for infant height estimation, where MobileNet, VGG16, GoogleNet, and AlexNet were chosen. Moreover, we investigated different MobileNet-based variants for infant weight estimation, including linear regression model, one-task model, and multi-task model. Several sets of experiments were carried out on the newborn dataset to validate the effectiveness of the proposed methods. The results show that the Mean Absolute Error (MAE) of different models are quite similar, with an average MAE < 1.1 cm and < 0.28 kg for height and weight estimation, respectively. Among them, the multi-task MobileNet has better temporal stability given its lower variance of measurement in a video.