Point2PartVolume: Human Body Volume Estimation From a Single Depth Image

Human body volume (BV) is a useful biometric feature for human identification and an important medical indicator for monitoring body health. Traditional BV estimation techniques such as underwater weighing and air displacement demand a lot of equipment and are difficult to be performed under some circumstances, for example, in clinical environments when dealing with bedridden patients. In this contribution, a novel vision-based method dubbed Point2PartVolume based on deep learning is proposed to rapidly and accurately predict the part-aware BVs from a single-depth image of the dressed body. First, a novel multitask neural network is proposed for jointly completing the partial body point clouds, predicting the body shape under clothing and semantically segmenting the reconstructed body into parts. Next, the estimated body segments are fed into the proposed volume regression network to estimate the partial volumes. A simple yet efficient two-step training strategy is proposed for improving the accuracy of volume prediction regressed from point clouds. Compared to existing methods, the proposed method addresses several major challenges in vision-based human BV estimation, including shape completion, pose estimation, body shape estimation under clothing, body segmentation, and volume regression from point clouds. Experimental results on both the synthetic data and public real-world data show that our method achieved average 90% volume prediction accuracy and outperformed the relevant state-of-the-art.