Body Part Labelling with Minkowski Networks

Abstract

Labelling body parts in depth images is useful for a wide variety of tasks. Many approaches use skeleton-based labelling, which is not robust when there is a partial view of the figure. In this work we show that Minkowski networks, which have recently been developed for 3D point cloud labelling of scenes, can be used to label point clouds with body part categories, achieving 85.6% accuracy with a full view of the figure, and 82.1% with partial views. These results are limited by a small sample size of our training data, but there is evidence that some of these ‘misclassifications’ may be correcting mistakes in the reference labelling. Overall, we demonstrate that Minkowski networks are effective for body part labelling in point clouds, and are robust to occlusion.