.png){kind=icon}
.png){kind=icon}
Return
6D object pose estimation based on dense convolutional object center voting with improved accuracy and efficiency
.png){kind=icon}
PRE
AI
DOI：10.1007/s00371-023-03113-4.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
6D object pose estimation is an important application of computer vision and a basic module in robotic manipulation, but dealing with occlusion in a cluttered environment, handling symmetries, and textureless surfaces, are real issues. Other issues with such systems are accuracy and efficiency. The recent two-stage methods perform well in terms of accuracy; however, a linear increase in their runtimes occurs due to the increase in the number of objects in a scene. This paper proposes a fully convolutional and parallel architecture that obtains the 3D translation and orientation for object poses from the same pixel-wise dense estimation. It exploits the same voting block for inliers for multiple instances, final 3D translation estimation, and quaternions aggregation. Only the center point estimation of the objects decreases the model's running time, while still useful for occlusions and texturelessness. Symmetries and varieties are handled with a loss function based on shape matching and the pose of the object. Our proposed approach has fewer parameters and takes less time to train and evaluate, achieving great accuracy. Experiments on LINEMOD and occlusion LINEMOD datasets using ADD (-S) and 2D projection evaluation metrics show that the proposed method outperforms state-of-the-art approaches for 6D pose estimation.
Keywords:
6D object pose estimation
Pixel-wise labeling
Pixel-wise center localization
Pixel-wise quaternion rotation