Prediction and Generation of 3D Functional Scene Based on Relation Graph

Functionality analysis of 3D shapes,as a means to understand and manipulate 3D environments,plays an important role in artificial intelligence.Human can predict the functionality of the object based on knowledge and experience without any surroundings.To this end,a method which uses relation graphs guide the generation of 3D functional scenes of a given object is proposed.The method consists of two steps:relation graph generation and scene generation.In the relation graph generation phase,a deep graph convolution generative network is used to predict a relation graph for the given object,which encodes the central and interacting objects as nodes and spatial relationships between objects as edges.In the scene generation phase,the generated graph and the given central object is taken as input to output the shapes and locations of interacting objects.Proposed method is tested on the interaction scene dataset for scene generation and compared with PG-DNN and PQ-NET.Chamfer distance and light field distance are used as evaluation indicators.Experimental results show that proposed method outperforms the state-of-the-art methods in functional scene generation and high-resolution functional scenes can be generated by using implicit representation of shapes.