Research And System Realization Of Occlusion Relationship Reasoning Based On Multi-level Feature Fusion

Reasoning the occlusion relationship of objects from monocular images is one of the basic tasks of computer vision.From the perspective of the observer,occlusion relationship reflects the relative depth difference between objects in the scene.Therefore,occlusion relation reasoning is also one of the important steps to perceive and understand the scene,which can be used for object detection,segmentation and 3D reconstruction.The occlusion relationship reasoning uses occlusion edges to describe objects,and orientation values to describe the occlusion order between objects.However,in the previous method,there are many false detections and missed detections of occlusion edge,and the false detection of occlusion relationship is of high rate.The reason is that they ignore the correlation and difference between the occlusion edge and the occlusion order,and their features are not fully extracted.In view of the above problems,this paper designs an occlusion relation inference system based on multi-feature fusion.The system inputs monocular pictures,outputs occlusion edge maps and occlusion relationship maps between objects,and it is divided into three parts:feature extraction module,semantic fusion module and stream prediction module.First of all,in the feature extraction module,for the problem of occlusion edge missed detection and discontinuity,this paper designs a simple and effective boundary refinement module to obtain spatial cues,and a boundary supervision module to achieve more accurate edge positioning.In order to obtain object-level occlusion edges,this topic uses a contour attention augmentation module to learn high-sense context features and enhance the contour perception of objects in the system.In order to solve the problem of inaccurate orientation value regression,this topic uses a multi-rate context learner to effectively perceive the information on both sides of the object contour.Next,this paper designs a semantic fusion module to solve the difference of multi-level structural features.In this module,the two subtasks of the occlusion relationship reasoning task,occlusion edge detection and occlusion orientation prediction,are separately semantically fused to obtain a unified adaptive feature map.Finally,in the stream prediction module,the occlusion edge and occlusion orientation are detected and predicted respectively.This module uses two kinds of loss functions to jointly supervise learning.The attention loss function solves the problem of extreme imbalance of positive and negative samples in occlusion edge detection,while the L1 loss function effectively handles the occlusion orientation continuous value prediction problem.The method in this paper validates and evaluates the method on the PIOD and BSDS ownership datasets from a quantitative and qualitative perspective.This method significantly improves the effect of occlusion relation inference while ensuring a low algorithm complexity.