Point Cloud Instance Segmentation With Semi-Supervised Bounding-Box Mining.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Deep learning has advanced point cloud instance segmentation but requires extensive, costly annotations.
• Current methods underutilize unlabeled or weakly labeled data, increasing annotation burden.
• There is a need for efficient point cloud segmentation methods that minimize data labeling requirements.

Purpose of the Study:

• To introduce the first semi-supervised point cloud instance segmentation framework (SPIB) that leverages both labeled and unlabeled bounding boxes.
• To reduce the dependency on dense, pixel-level annotations in point cloud instance segmentation.
• To enable competitive segmentation performance with significantly lower annotation effort.

Main Methods:

• A two-stage learning procedure involving bounding box proposal generation and instance mask mining.
• Semi-supervised training with perturbation consistency regularization (SPCR) for self-supervision.
• Novel semantic propagation, property consistency graph, and occupancy ratio guided refinement modules for mask generation and refinement.

Main Results:

• The SPIB framework demonstrates competitive performance against state-of-the-art fully-supervised methods on the ScanNet v2 dataset.
• The proposed semi-supervised approach effectively utilizes unlabeled bounding box data.
• SPCR provides robust self-supervision for bounding box proposal generation.

Conclusions:

• SPIB presents a viable and efficient solution for point cloud instance segmentation with reduced annotation costs.
• The framework's ability to use unlabeled bounding boxes opens new avenues for leveraging large unlabeled datasets.
• This work significantly contributes to making deep learning-based point cloud segmentation more accessible and practical.