Scene Understanding with Support Relation Inference for Exoskeleton Robot

The exoskeleton robot in the rapid development requires physical relation interpretation about scene understanding to fulfill effective interaction. Although traditional scene understanding provides good performance of object recognition and semantic segmentation, it tells us little about physical relation interpretation. Meanwhile, there are few methods to demonstrate the contextual relations among objects in the scene. While some of them artificially define prior constrains and solve the equation by minimizing the objective function, others aim at perceiving the stability of 3D blocks to classify the support relation to several classes. In this paper, a novel approach is proposed which can automatically get the prior knowledge based on deep learning and Conditional Random Fields(CRF) to infer support relationship. In addition, unlike all of previous methods applying the post-processing step after segmenting or recognizing objects, our method is based on end-to-end deep network which could be trained by back-propagation algorithm. The support relationship can be directly obtained as the output from a RGB picture input through our approach, where no other intermediate steps are needed. We show the effectiveness of the proposed method on the NYU-Depth V2 dataset in which we manually re-annotate 1449 images.