Scalable SoftGroup for 3D Instance Segmentation on Point Clouds.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Current 3D instance segmentation methods rely on hard semantic predictions, leading to error propagation and poor results.
• Existing fast methods are not suitable for real-time applications due to computational bottlenecks, particularly in large-scale scenes.

Purpose of the Study:

• To develop an accurate and scalable 3D instance segmentation network.
• To address limitations of existing methods, including error propagation and computational inefficiency.

Main Methods:

• SoftGroup allows points to associate with multiple classes, mitigating semantic prediction uncertainty.
• SoftGroup++ optimizes the k-Nearest Neighbor (k-NN) module using octree k-NN, class-aware pyramid scaling, and late devoxelization for improved scalability.
• The methods suppress false positives by learning to categorize them as background.

Main Results:

• SoftGroup and SoftGroup++ surpass state-of-the-art baselines by 6%-16% in AP50.
• SoftGroup++ achieves an average 6x speedup on large-scale scenes compared to SoftGroup.
• The SoftGroup framework demonstrates versatility, improving object detection and panoptic segmentation.

Conclusions:

• SoftGroup provides a robust solution for accurate 3D instance segmentation by handling semantic uncertainty.
• SoftGroup++ significantly enhances scalability, making real-time 3D instance segmentation feasible for large-scale environments.
• The proposed methods offer a general framework applicable to various 3D computer vision tasks.