Multi-Head Multi-Scale Feature Fusion Network for Semantic Scene Completion

Abstract

In recent years, 3D semantic scene completion has achieved rich development and has been widely used. Part of the work is to study the fusion mode of multimodal features under a basic completion framework in order to obtain better geometric or semantic features. Part of the work is to study the improvement of the completion framework under the condition of using a single depth or RGB-D multimodal input. The semantic scene completion network proposed by us is based on the basic framework, adds attention mechanism, and creatively uses multi-scale method to improve the network, so as to obtain better 3D semantic scene completion effect.