The Research Of Relation-based Indoor 3D Object Detection

In recent years,with the development of intelligent and unmanned technologies,typ-ical applications of 3D vision such as autonomous driving,intelligent robots and intelli-gent logistics have attracted lots of attention.3D object detection acts as the ”intelligent eye” in 3D vision,and is the basis of many 3D vision tasks.Based on 3D reconstruction,indoor 3D object detection is aimed to obtain information such as the position,shape,ori-entation and category of objects in the scene.This saves human resources and achieves automation by localizing and recognizing the target objects accurately.However,in our real world,objects belonging to the same category may vary in shape and appearance.3D object detection mainly performed object detection based on the geometric features of point clouds in the past,ignoring the inherent relation contexts between objects.In fact,there are rich relation contexts between objects in indoor scenes,and how to use these relations that imply a large amount of information to help 3D object detection remains a challenge.Recent methods mainly use context encoding or graph structures to extract implicit relational information,which is inefficient and unstable.In addition,there are many types of objects in indoor scenes,and the relation contexts between objects will inevitably contain redundant or even invalid information.Directly using the information without filtering may lead to ambiguity while detection.How to extract accurate relation contexts from indoor scenes with complex layouts is also a challenge.In response to the above challenges,this paper proposes effective solutions,includ-ing the typical semantic and spatial relationship recognition algorithms of indoor scenes,which solves the challenge of effective relation reasoning between objects.This paper designs and implements an attention-based relation module based on the attention mech-anism,which can extract accurate,effective and robust relation contexts.This eliminates the negative influences of confusing relational information,and helps achieve more accu-rate 3D object detection.The main innovations of this article are summarized as follows:1.Semantic and spatial relation reasoning: Inspired by the typical relationships between objects in indoor scenes,this paper designs semantic and spatial relations in-cluding the same category,the same instance,support and hang on relations.This paper proposes a lightweight but effective relation recognition module that can effectively ex-tract the relation contexts as well as relation features between objects,which enrichs the feature expression in object detection and also promote the ability of scene understanding.2.Relation attention mechanism based on Transformer: Inspired by Trans-former,this paper proposes a relation module based on the attention mechanism.Com-bined with the objectness module and relation reasoning,we can model the importance of different relation contexts,and thus make full use of the beneficial information in relation contexts,avoiding the negative influence of redundant information.In this way,it further improves the accuracy and robustness of 3D object detection.This paper applies the relation module to the popular object detection frameworks,constructing a complete 3D object detection system.Numerous experiments are carried out on three challenging datasets.Experimental results prove the effectiveness and uni-versality of this method.This paper also applies the plug-and-play relation module to the highly versatile MMDetection3 D platform,which improves the reusability of our method and promotes related research in the field.