Category-level Object Pose Estimation

With the changes of life in recent years,the applications of AR,robotics and other technologies have become more and more popular,object pose estimation becomes an important problem that needs to be solved.Due to object pose has 6 degrees of freedom（3 rotations and 3 offsets）,object pose estimation problem is also called object 6D pose estimation problem.This problem aims to estimate the position and direction of the object in the camera coordinate system.Solves object pose estimation problems can propose more effective solutions for various problems related to scene understanding,AR,robot control and navigation.In recent years,with the development of depth sensors,the object 6D pose estimation problem has become more and more popular.Existing object pose estimation solutions mainly focus on instance-level object 6D pose estimation problem.In the settings of instance-level object 6D pose estimation problem,the standard CAD models corresponding to objects are known in the test stage,so the problem is easier.But at the same time it also limits that these solutions cannot be applied to the scenes where objects that never been seen and no known CAD models.However,this condition is necessary for the object pose estimation algorithm is applied in daily life.Recently,there are some solutions focus on category-level object pose estimation problem,but these works aim to first reconstruct the CAD model corresponding to objects,and then estimate the current pose of the object between current point cloud and reconstructed CAD model.Existing studies mainly focused on how to better reconstruct the CAD model corresponding to the object.However,the reconstruction task of the 3D model is a difficult task,and the process of reconstruct the CAD model is time-consuming,which has led to some jobs still have shortcomings in terms of accuracy,speed,and cannot meet application requirements.Based on the shortcomings of the above work,we propose a novel category-level object pose estimation solution named KeyPoseNet.Our method removes the dependence on the reconstructed CAD model of the object,and uses keypoints as a proxy for the object’s pose.At the same time,a memory module is proposed,the memory module selects the features in the standard space closest to the current object,and then generates the keypoints of the standard model,which can be used to solve the key problem of category-level object pose estimation—the lack of corresponding CAD model in the test stage.Then,the three-dimensional point set registration algorithm is used to estimate pose.At the same time,due to the flexibility of the memory module,we can also unify single-frame object pose estimation and multi-frame object pose tracking tasks.Experiments proves that our method does not need to reconstruct the corresponding CAD model of the object,and the final result can still be comparable to the state-of-the-art method.The main contributions of this paper are:1.Propose a novel class-level object pose estimation method named KeyPoseNet.Compared with existing work,KeyPoseNet does’t model dense standard model（CAD model）of the target object,but a sparse keypoint set.This method avoids the complex task of modeling CAD models while maintaining high accuracy of pose estimation.2.Based on KeyPoseNet,a memory module is proposed.This module unifies the category-level object pose estimation problem based on a single frame and the object pose tracking task based on continuous frames.Different tasks can be solved using the same network without additional adjustments.3.The original NOCS data set is expanded,and the data and annotations in the standard space required by KeyPoseNet training framework are added to provide support for the follow object pose estimation researches based on sparse models such as keypoint sets.