SCN: Switchable Context Network for Semantic Segmentation of RGB-D Images

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Context representations are crucial for semantic image segmentation.
• Depth data offers geometric information often missing in RGB images.
• Optimizing Convolutional Neural Networks (CNNs) with depth data for enhanced segmentation accuracy is challenging.

Purpose of the Study:

• To present a novel Switchable Context Network (SCN) for semantic segmentation of RGB-D images.
• To effectively utilize depth data for constructing more discriminating context representations.
• To improve the accuracy and coherence of semantic segmentation networks.

Main Methods:

• The SCN utilizes depth data to identify objects across multiple image regions.
• It analyzes image region information to discern different characteristics.
• Network branches are selectively employed through a switching mechanism based on extracted features.

Main Results:

• The SCN generates context representations aware of image structures and object relationships.
• This leads to more coherent learning in semantic segmentation.
• The proposed SCN demonstrates superior performance compared to state-of-the-art methods on public datasets.

Conclusions:

• The Switchable Context Network (SCN) effectively integrates depth data for enhanced semantic segmentation.
• The network's ability to switch branches based on image characteristics improves context representation.
• SCN offers a promising approach for accurate RGB-D semantic segmentation.