Refining Weights for Enhanced Object Similarity in Multi-perspective 6Dof Pose Estimation and 3D Object Detection

At the moment, there are increasing trends in using deep learning for 6 Dof pose estimation and 3D object detection. Recognising objects and determining their 3D positioning and orientation in the scene has numerous applications in robotics. However, due to the variety of objects within the real world, the challenge is complicated. They have various 3D shapes, and their appearance on images can be affected by lighting, clutter in the scenery, and occlusions between objects. Matching feature points between 3D models and images are used to solve the challenge of 6-Dof object pose estimation. We propose a hybrid method for 6 Dof and 3D object detection using a modification of the Resnet18 pre-trained model, combined with object matching and the object symmetric method. We conduct experimental tests with video files and a live view cam. Visualisation test variations are carried out on an object in a single perspective view, a single object in a multi-perspective view, a multi-object in a single perspective view, and a multi-object in a multi-perspective view. We evaluated the performance of this hybrid method on 15 types of objects from Linemod (LM) and Linemod Occluded (LM-O) dataset, resulting in a total training loss of 0.0533049 and a total validation loss of 0.0481146.