Relative Pose Estimation between Image Object and ShapeNet CAD Model for Automatic 4-DoF Annotation

Estimating the three-dimensional (3D) pose of real objects using only a single RGB image is an interesting and difficult topic. This study proposes new pipeline to estimate represent object in with 4-DoF annotation matching CAD model. The proposed method retrieves candidates from ShapeNet dataset utilizes pose-constrained 2D renderings find best estimation consists several steps learned networks followed by similarity measurements. First, image, category region are determined segmented. Second, 3-DoF rotational estimated pose-contrast network segmented region. Thus, rendering images generated based on result. Finally, measurement performed model determine 1-DoF focal length camera align object. Conventional methods employ 9-DoF parameters due unknown scale both However, this shows that between enough facilitates projection space for image-graphic applications such as Extended Reality. In experiments, performance analyzed ground truth comparing triplet-loss learning method.