.png){kind=icon}
.png){kind=icon}
Return
Residual Regression Pose Estimation Network Based on Dynamic Reference Points
.png){kind=icon}
PRE
AI
DOI：10.1109/JSEN.2026.3678960.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
To address challenges such as specular reflections on metal surfaces, limited texture, and severe occlusion in industrial scenarios, this article proposes a residual regression pose estimation network based on dynamic reference points, which achieves robust six-degree-of-freedom (6DoF) pose prediction of workpieces from a single RGB-D image. Our method first performs preprocessing, including cropping and normalization of the RGB-D image according to object detection results, to improve input data quality. A dynamic reference point allocation strategy, designed based on object size and curvature information, is combined with a residual representation of 3-D object coordinates, effectively reducing the output space and enhancing attention to key geometric regions. The network introduces a bidirectional fusion module that integrates RGB and depth information to extract and combine appearance and geometric features in a multiscale manner. Finally, by establishing dense correspondences, the network jointly predicts the 6DoF object pose. On the LM-Occluded dataset, the proposed method achieves an average improvement of 4.8% in the ADD-S 0.1d metric over existing baselines, achieves a 7% average improvement on our custom metal workpiece dataset, and maintains high robustness under complex occlusion conditions, demonstrating the effectiveness of our approach.
Keywords:
Industrial workpieces
point cloud
RGB-D fusion
six-degree-of-freedom (6DoF) pose estimation