Positive-Negative Receptive Field Reasoning for Omni-Supervised 3D Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Neural networks often struggle with 3D segmentation due to limited supervision on output predictions.
• Informative feature representation is crucial for accurate 3D segmentation tasks.

Purpose of the Study:

• To introduce the first omni-scale supervision method for 3D segmentation.
• To improve feature learning in neural networks for enhanced 3D segmentation performance.

Main Methods:

• Proposed a gradual Receptive Field Component Reasoning (RFCR) method with Receptive Field Component Codes (RFCCs) to supervise intermediate layers.
• Introduced RFCR-NL model with Negative RFCCs (NRFCCs) for enhanced supervision via negative learning.
• Developed Feature Densification with centrifugal potential to create more unambiguous features, akin to entropy regularization.

Main Results:

• Significantly improved performance across three datasets for both fully and weakly supervised 3D segmentation tasks.
• Achieved competitive results when integrated into three prevailing neural network backbones.
• Demonstrated the effectiveness of omni-scale supervision in unlocking the potential of feature learning.

Conclusions:

• The proposed omni-scale supervision method effectively addresses limitations in 3D segmentation.
• RFCR and Feature Densification contribute to more robust and informative feature representations.
• The method shows broad applicability and strong performance across various segmentation scenarios.