Instance Consistency Regularization for Semi-Supervised 3D Instance Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Analysis

Background:

• 3D instance segmentation requires extensive labeled data, which is costly to acquire.
• Existing semi-supervised methods often struggle with noisy pseudo-labels from semantic information.
• Leveraging unlabeled data is crucial for advancing 3D instance segmentation.

Purpose of the Study:

• To develop a semi-supervised 3D instance segmentation method that relies solely on instance consistency regularization.
• To mitigate the noise and collapse issues inherent in semantic pseudo-labels within self-training frameworks.
• To improve the efficiency and accuracy of 3D instance segmentation using unlabeled data.

Main Methods:

• Proposed InsTeacher3D, a novel self-training network for semi-supervised 3D instance segmentation.
• Introduced DKNet, a base model distinguishing instances via discriminative instance kernels without semantic reliance.
• Developed a new instance consistency regularization framework to generate and utilize high-quality instance pseudo-labels.

Main Results:

• InsTeacher3D demonstrated significant performance improvements over state-of-the-art semi-supervised methods.
• The method effectively leverages unlabeled data through pure instance knowledge.
• Experimental results validated the approach on multiple large-scale datasets.

Conclusions:

• Solely relying on instance consistency regularization is a viable and effective strategy for semi-supervised 3D instance segmentation.
• InsTeacher3D offers a robust solution to the challenges posed by noisy semantic pseudo-labels.
• The proposed method advances the field by enabling better utilization of unlabeled data for 3D instance segmentation.