.png){kind=icon}
.png){kind=icon}
Return
Active learning with joint probabilistic modeling for point cloud semantic segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.114171.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
With advancements in sensing technologies, the demand for point cloud semantic segmentation has grown significantly across various applications, while current deep learning-based methods rely heavily on costly, well-annotated datasets. Recently, label-efficient learning strategies have been explored to reduce annotation demands, with active learning emerging as a preferred approach by selectively annotating only the most informative samples. However, existing point cloud active learning methods often depend solely on neural network softmax scores for sample selection, which can introduce bias and be affected by overconfidence in network predictions. To overcome this limitation, we propose an active learning framework with Joint Probabilistic modeling (JoPro), aiming to select unlabeled points that can provide more post-annotation information. At the core of JoPro is a novel probabilistic model that efficiently captures the distribution of embedded features to generate richer probabilistic representations for unlabeled data. Utilizing this probabilistic modeling, we propose a feature mixing stability metric to identify uncertain points near decision boundaries, ensuring more informative sample selection. Furthermore, a cluster-aware hybrid contrastive regularization method is incorporated to maximize the utilization of unlabeled data to enhance training of the segmentation model. Our proposed active learning framework achieves competitive results on popular benchmarks, delivering near fully supervised performance with only 1 % of the annotation budget.
Keywords:
point cloud semantic segmentation
active learning
probabilistic modeling
label-efficient learning
feature mixing stability
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W