.png){kind=icon}
.png){kind=icon}
Return
Weakly supervised point cloud semantic segmentation using pseudo-label reliability and consistency regularization
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2025.130241.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Weakly supervised point cloud semantic segmentation using sparse 3D labels effectively reduces the high cost associated with annotating dense 3D labels. This paper explores pseudo-labeling and consistency regularization techniques in weakly supervised point cloud semantic segmentation, and proposes a consistency regularization sparse convolutional network based on pseudo-label reliability. We identify issues with pseudo-labels selected based on confidence, as high confidence but incorrect labels are selected, while low confidence but potentially useful labels are discarded. To address these challenges, we combine confidence and uncertainty to measure the reliability of pseudo-labels and apply different consistency constraints based on this reliability. We also employ entropy regularization loss to more accurately distinguish pseudo-label reliability. Considering the significant impact of training strategies on model performance, we investigate the influence of augmented point cloud and distribution alignment loss, both associated with consistency regularization, on model training. Extensive experiments conducted on two most popular large-scale benchmark datasets, S3DIS and ScanNet, demonstrate the effectiveness of our proposed method. The experimental results show that our method achieves optimal performance in weakly supervised point cloud semantic segmentation.
Keywords:
Weakly supervised
Point cloud
Semantic segmentation
Pseudo-label
Consistency regularization
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W