To assist frontline health workers identify underweight neonates and monitor their growth, we are developing a smartphone-based technology that provides accurate, timely, geo-tagged and tamper-proof weight estimation.
Our AI solution for estimating the weight of a baby learns to map a set of baby pictures or a video to the baby’s weight. Currently, we are exploring deep learning methods that take as input a set of video frames,and directly estimate the baby’s weight by performing implicit reasoning about the baby’s shape and volume. In the past, we have also explored 3D approaches that leverage a parametric, deformable 3D model to fit the baby’s appearance, followed by volume and weight estimation using a constant density. To this end, we are creating a 3D parametric model of Indian babies.
To obtain a better understanding of baby shape, the data presented to the model consists of video that captures the baby from a range of angles by sweeping a smartphone directly above a baby. Ground-truth for the AI system is obtained using accurate digital weighing scales, both in rural home and hospital settings. The approach also relies on having an object of standard dimensions placed next to the baby. This object serves as a reference to translate the scale of the baby from the image (pixels) space to a metric space, and is an important input to weight estimation.