Monocular Object Pose Estimation Using Specific Object Imaging Techniques
Shuqi Luo, Jun Li, Yaojin Xie, Rao Zhang · 2024
Estimating robust 6D object poses using RGB images under conditions of clutter or occlusion remains a challenging task. Existing methods typically either extract features from the entire scene or apply deep networks for feature extraction following object detection and semantic segmentation. However, these approaches can lead to incorrect learning, causing models to fail in accurately capturing the core characteristics of objects and resulting in errors when matching 2D images with 3D models. Furthermore, in complex and dynamic scenes, the detection performance of current methods still requires improvement, underscoring the need for further research and algorithm optimization. To address this issue, we propose a 6D pose estimation method based on specific object imaging. Our method actively and intelligently extracts image information of specific targets through object imaging, thereby avoiding global processing of the entire scene and significantly reducing the amount of irrelevant information processed. This enables the model to more accurately learn the complete feature information of target objects, achieving higher accuracy in complex and dynamic application scenarios. Experiments on the public LM-O dataset demonstrate that our method achieves significant improvements in the ADD(-S) metric compared to existing techniques.