Point Cloud Segmentation Using Sparse Temporal Local Attention

However, Point clouds are a key modality used for perception despite a number of successful recent approaches exploiting in autonomous vehicles, providing the means sequential data from 2D video streams for improved for a robust geometric understanding of the surrounding segmentation performance [Hu et al., 2020a; Li et al., 2018; environment. However despite the sensor Paul et al., 2020; Zhu et al., 2019; Jain et al., 2019], there outputs from autonomous vehicles being naturally has been limited exploration into leveraging temporal priors temporal in nature, there is still limited exploration for point cloud segmentation. Existing approaches either calculate of exploiting point cloud sequences for 3D semantic strict correspondences between point features across segmentation. In this paper we propose a novel frames [Cao et al., 2020] or perform global attention [Shi Sparse Temporal Local Attention (STELA) module et al., 2020] between whole point clouds. In the case of the which aggregates intermediate features from a local former, a breakdown of nearest-point matching due to displacement neighbourhood in previous point cloud frames between adjacent point clouds can result in the to provide a rich temporal context to the decoder.