Box-Supervised 3D Instance Segmentation With Level Set Evolution and Cross-View Consistency.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• Weakly supervised 3D instance segmentation reduces annotation costs compared to fully supervised methods.
• Box annotations offer a balance between labeling efficiency and supervision strength.
• Existing box-supervised methods often use iterative pseudo-labeling, which is sensitive to label quality.

Purpose of the Study:

• To develop an end-to-end framework for 3D instance segmentation using only box annotations.
• To eliminate the need for explicit pseudo-label generation and iterative retraining.
• To improve segmentation accuracy and training stability in weakly supervised settings.

Main Methods:

• A boundary-aware refinement module using level set evolution to learn instance boundaries from boxes.
• A multi-scale geometric augmentation module with cross-view consistency constraints to handle overlapping regions.
• A multi-objective optimization framework to enhance training stability and performance.

Main Results:

• The proposed method achieves state-of-the-art (SOTA) performance on multiple indoor and outdoor benchmark datasets.
• The framework directly learns instance masks from box annotations without iterative pseudo-labeling.
• Performance closely approaches that of fully supervised methods.

Conclusions:

• The end-to-end framework effectively utilizes box annotations for 3D instance segmentation.
• The novel modules improve boundary learning, semantic ambiguity resolution, and training stability.
• This approach offers a more efficient and robust alternative to existing weakly supervised methods.