Point Cloud Instance Segmentation with Inaccurate Bounding-Box Annotations.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Point cloud instance segmentation demands precise, dense point-level annotations, which are costly and time-consuming.
• Existing methods struggle with incomplete or inexact supervision, particularly in complex 3D environments.
• Inaccurate supervision severely impacts the generalization performance of deep learning networks.

Purpose of the Study:

• To develop the first weakly supervised point cloud instance segmentation framework utilizing inaccurate box-level labels.
• To enhance the generalization ability of deep networks when trained with noisy bounding-box annotations.
• To address the under-explored challenge of inaccurate supervision in 3D point cloud analysis.

Main Methods:

• A novel self-distillation architecture is proposed to leverage cheap but noisy bounding-box annotations.
• Consistency regularization is employed to distill self-knowledge from data perturbations and historical predictions.
• A progressive sample selection and label correction mechanism based on historical consistency is implemented.

Main Results:

• The proposed framework effectively boosts generalization performance despite inaccurate supervision.
• The method demonstrates robustness in handling inexact and inaccurate annotations on the ScanNet-v2 dataset.
• Self-distillation and consistency regularization prevent the network from overfitting noisy labels.

Conclusions:

• The developed framework offers a viable solution for point cloud instance segmentation with reduced labeling effort.
• This approach significantly improves the practical applicability of deep networks in real-world 3D scenarios.
• The method validates the effectiveness of self-distillation and label correction for weakly supervised learning with noisy data.