3D-scene modelling of professional gestures when interacting with moving, deformable and revolving objects

In this paper, we present good practices of applying and extending Random Decision Forests (RDFs) for the 3D modelling of scenes where humans interact with moving, deformable and revolving objects in a professional context. We apply our method to two use-cases; the first is in the industrial context of the luxury leather good production while the second is in an atelier specialised in the wheel-throwing art of pottery. In the first use-case we use a single RDF, while for the second one of pottery, we extend the typical application of RDFs, by introducing the Hierarchical Random Decision Forests (HRDFs). More precisely, we use three RDFs in a tree structure architecture. The parent RDF is used to create a rough initial segmentation of the scene, while the two children RDFs are used to further classify the regions of the left and right arm, hand and fingers respectively. Results demonstrate that the proposed algorithm is sufficient for the accurate classification of scenes where humans interact with objects by using hand gestures in both simple and complex scenarios.