Design And Implementation Of Automatic Labeling System For Rigid Object 6D Pose Estimation

In recent years,rigid-object 6D pose estimation has attracted much attention due to its wide application in fields such as augmented reality,autonomous driving and robot manipulation.With the development of artificial intelligence,deep learning-based methods are becoming more effective,and the performance of deep networks relies heavily on a large number of accurately labeled datasets.Traditional labeling methods rely on manual completion by annotators,but such methods suffer from inefficiency,high cost,and unstable quality.With the increase of data demand,semi-automatic labeling methods based on depth sensors have emerged,but these methods also require a certain amount of manual labeling and complex manual operations to complete the semiautomatic labeling process,and these methods usually use flat calibration plates to assist in positioning,which are prone to problems such as occlusion and reflection.To address these problems,this paper proposes an automatic labeling scheme for rigid 6D pose estimation,which achieves automatic labeling through the chain pose transformation relationship between the object,camera,and labeling device.Since the positioning effect of the plane calibration plate is easily affected by the shooting angle and light factor,a three-sided marking table is designed as an auxiliary marking device.The automatic labeling scheme mainly includes two parts: preparation stage and automatic labeling stage.To complete the preparation phase,this paper first proposes a multiple-task pixel-level voting network(Multiple Task-PVNet)prediction method to achieve robust positioning of the three-sided labeling table and accurate prediction of the target object mask,and the experimental results show that the model has high accuracy in predicting both the position of the three-sided labeling table and the object mask.Secondly,this paper proposes a model filtering strategy based on the object mask to construct local point clouds of the object by filtering the scene model.Then,the initial pose of the object is obtained by filtering the results of point cloud registration using the point cloud registration and screening strategy proposed in this paper.Finally,this paper combines the object mask with the area-based pose refinement method to realize the correction and optimization of the initial pose.At this point,the fixed position relationship between the object and the three-sided labeling table can be obtained,and the preparation stage is completed.Then,the automatic labeling of object poses can be realized in the automatic labeling phase by the chain transformation rule.In order to evaluate the effectiveness of the above automatic labeling scheme,this paper uses the existing mature object pose estimation algorithm PVNet to verify the accuracy of labeling.The experimental results show that the automatic labeling scheme in this paper can greatly reduce the labeling cost while maintaining the labeling quality.In the case of the same number of labeled images,the automatically generated training data achieves better object pose estimation accuracy than the manually labeled data;when the number of automatically labeled images reaches about 2000,the accuracy of the pose estimation algorithm for most objects is close to saturation.Based on the automatic labeling scheme,this paper develops an automatic labeling system for rigid body 6D pose estimation.The system visualizes,systematizes,and processes the automatic labeling scheme through the functional modules of data acquisition,computation of fixed transformation,and automatic labeling to achieve automatic labeling of object pose in the image to be labeled.The evaluation of the results of the system shows that the poses generated using the automatic labeling system have good accuracy.