SegGroup: Seg-Level Supervision for 3D Instance and Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Current 3D point cloud segmentation methods require extensive point-level annotations, leading to high costs.
• Weakly-supervised approaches are needed to reduce the annotation burden for 3D scene understanding.

Purpose of the Study:

• To develop an efficient weakly-supervised method for point cloud instance and semantic segmentation.
• To leverage instance locations as a minimal annotation signal.

Main Methods:

• A novel weakly-supervised approach utilizing single-point location annotations per instance.
• Over-segmentation pre-processing to convert point annotations to segment-level labels.
• A Segment Grouping Network (SegGroup) to generate point-level pseudo-labels from seg-level annotations.

Main Results:

• The proposed seg-level supervised method (SegGroup) achieves performance comparable to fully supervised methods.
• SegGroup outperforms existing weakly-supervised methods under a fixed annotation budget.
• The method effectively reduces annotation effort for 3D point cloud segmentation tasks.

Conclusions:

• Leveraging instance locations is an effective strategy for weakly-supervised 3D point cloud segmentation.
• The SegGroup network provides a viable solution for training segmentation models with minimal supervision.
• This approach significantly lowers the barrier to entry for applying deep learning to 3D scene segmentation.