Equivariant Spatio-Temporal Self-Supervision for LiDAR Object Detection

arXiv:2404.11737v1 Announce Type: new
Abstract: Popular representation learning methods encourage feature invariance under transformations applied at the input. However, in 3D perception tasks like object localization and segmentation, outputs are naturally equivariant to some transformations, such as rotation. Using pre-training loss functions that encourage equivariance of features under certain transformations provides a strong self-supervision signal while also retaining information of geometric relationships between transformed feature representations. This can enable improved performance in downstream tasks that are equivariant to such transformations. …

abstract arxiv cs.cv detection feature features functions however lidar localization loss object perception popular pre-training representation representation learning rotation segmentation supervision tasks temporal training training loss type