DSS6D:Dual Stream Fusion Network for 6D Pose Regression of Symmetric Objects

This paper proposes a novel framework for 6D pose estimation. In the case of symmetrical objects, objects may have multiple ground-truth poses, this one-to-many relationship may lead to estimation errors. To address this problem, we introduce a symmetry-invariant pose distance metric called the mean (maximum) grouped primitive distance or a (M)GPD. The A(M)GPD loss enables the model to converge to the correct pose so that all minima on the A(M)GPD loss can be mapped to the correct pose. We use a fully convolutional network to extract point-wise features from RGB-D data. By driving the connection between features through an attention mechanism, regress the 6D pose without post-refinement. Compared with other frameworks, our framework is more suitable for the 6D pose of symmetrical objects. We conduct experiments on the YCB-Video and TLESS datasets, and the results demonstrate that the model outperforms the state-of-the-art methods when dealing with symmetrical objects, demonstrating that the framework has high accuracy and low computational cost.