On the Estimation of Children's Poses

Deep Learning architectures have obtained significant results for human pose estimation in the last years. Studies of the state of the art usually focus their attention on the estimation of the human pose of adults people depicted in images. The estimation of the pose of child (infants, toddlers, children) is sparsely studied despite it can be very useful in different application domains, such as Assistive Computer Vision (e.g. for early detection of autism spectrum disorder). The monitoring of the pose of a child over time could reveal important information especially during clinical trials. Human pose estimation methods have been benchmarked on a variety of challenging conditions, but studies to highlight performance specifically on children's poses are still missing. Infants, toddlers and children are not only smaller than adults, but also significantly different in anatomical proportions. Also, in assistive context, the unusual poses assumed by children can be very challenging to infer. The objective of the study in this paper is to compare different state of art approaches for human pose estimation on a benchmark dataset useful to understand their performances when subjects are children. Results reveal that accuracy of the state of art methods drop significantly, opening new challenges for the research community.