Unsupervised 3D Object Segmentation of Point Clouds by Geometry Consistency.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• 3D object segmentation from point clouds typically requires extensive human annotations.
• Existing methods lack unsupervised approaches for simultaneous multi-object identification.

Purpose of the Study:

• To propose the first unsupervised method (OGC) for 3D object segmentation from raw point clouds.
• To automatically discover rigid objects without human annotations by leveraging motion patterns.

Main Methods:

• Developed an unsupervised method (OGC) comprising an object segmentation network, a self-supervised scene flow estimator, and an object geometry consistency component.
• Utilized dynamic motion patterns in sequential point clouds as supervision signals.
• Designed loss functions to enforce multi-object rigid consistency and object shape invariance.

Main Results:

• Demonstrated superior performance in object part instance segmentation and general object segmentation across five datasets.
• Achieved effective 3D object discovery in both indoor and challenging outdoor scenarios.
• Validated the method's ability to discover object geometry without annotations.

Conclusions:

• The proposed OGC method offers a novel unsupervised approach for 3D object segmentation.
• Leveraging motion patterns provides effective supervision for discovering object geometry.
• The method shows strong generalization capabilities across diverse environments and segmentation tasks.