Footnote:
Diese Datenquelle enthält auch Bestandsnachweise, die nicht zu einem Volltext führen.
Description:
Densely annotating LiDAR point clouds remains too expensive and time-consuming. However, large-scale datasets are crucial for robustness in dense 3D perception tasks. In this thesis, we tackle data-efficient LiDAR semantic segmentation, with the goal of reducing the cost of labeling while retaining performance. First, we propose using line-scribbles to annotate LiDAR point clouds and release ScribbleKITTI, the first scribble-annotated dataset for LiDAR semantic segmentation. We further present a pipeline to close the performance gap when using such weak annotations. Our pipeline comprises three stand-alone contributions that can be combined with any LiDAR semantic segmentation model to significantly improve performance while introducing no additional computational/memory costs during inference. However, despite line-scribbles' efficacy, keeping pace with the sheer volume of data is still difficult. Labeling must therefore be done selectively. To this end, we explore active learning (AL) for LiDAR semantic segmentation while considering common labeling techniques such as sequential labeling to iteratively and intelligently label a dataset under a fixed budget. We propose a discwise approach (DiAL), where in each iteration, we query the region a single frame covers on global coordinates, labeling all frames simultaneously and resolving the two major challenges that emerge upon this choice. Finally, we tackle the weaknesses of weakly- and semi-supervised LiDAR semantic segmentation models by improving boundary estimation and reducing the high false negative rates for small objects and distant sparse regions. We construct an image-guidance network (IGNet) that distills high-level feature information from a domain-adapted synthetically trained 2D semantic segmentation model to improve 3D performance while not introducing additional annotation costs. Our final model achieves 98% relative performance to fully-supervised training while only using 8% labeled points.