Research On The Object-level Depth Ordering

In computer vision,depth ordering is an essential yet challenging problem describing the layer relationship between regions,which is fundamental in the image analysis and understanding and can be further applied in many high-level visual tasks,e.g.,image and video coding,autopilot,motion tracking and analysis,3D scene reconstruction.Enlightened by the research status of depth ordering,this paper proposes an object-level depth ordering approach,which simulates the human visual mechanism and achieves good performance and low complexity.The main works of this paper include three parts:occlusion edge detection,object proposal and depth ordering.(1)Occlusion edge detection.This paper introduces the process of occlusion edge detection,including edge sample construction(image partition and feature extraction),feature selection,sample reconstruction and classifier learning.Above all,a supervised sparse regression is utilized for more reasonable feature selection,and sparse coding with Huber loss is designed to reconstruct samples.Then,the kernel ridge regression is used for training.Finally,the quantitative and qualitative experiments are done on the D-order dataset and NYU2 dataset to demonstrate the presented approach.(2)Object proposal.This paper introduces the object surface cue into proposals,which is rarely discussed in the existing methods.Based on the sliding window mechanism,each sampled window obtains its objectness score with the detected occlusion edge map,measuring the likelihood of containing an object.Moreover,edge response fusion and adaptive normalization criterion make occlusion response more reliable,leading to the promotion of score ranking.Finally,the experiments are done on the PASCAL VOC 2007 dataset and MS COCO 2014 dataset.(3)Depth ordering.For local inference,a triple descriptor is defined to combine effective junction cue with convexity cue to judge local depth order jointly.Then a weighted directed graph model is introduced to correct local results globally.Similarly,the experiments are done on the D-order dataset and NYU2 dataset to evaluate our approach.The essential purpose of this paper is to infer the depth order in a monocular image,and it improves occlusion edge detection,object proposal and depth ordering jointly to boost the performance of depth order in a full image.