The article introduces a novel coarse-to-fine approach for 3D instance segmentation, named Spherical Mask. It is based on spherical representation, which overcomes limitations of existing methods, such as instance size overestimation and false negative error accumulation. The method proposes to estimate each instance with a 3D polygon using center and radial distance predictions to avoid excessive size estimation. It also introduces two margin-based losses for point migration to enforce corrections for false positives/negatives. Experimental results show that this method outperforms existing works, demonstrating the effectiveness of the new instance representation with spherical coordinates.
Publication date: 18 Dec 2023
Project Page: https://arxiv.org/abs/2312.11269
Paper: https://arxiv.org/pdf/2312.11269
