Towards Achieving Robust Low-level and High-level Scene Parsing.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Scene segmentation is a core computer vision task requiring detailed spatial and contextual information.
• Existing methods using "dilation" and "skip" connections in Convolutional Neural Networks (CNNs) have limitations in preserving fine-grained details.
• Optimizing parameterization of skip layers and incorporating low-level features are crucial for enhanced parsing performance.

Purpose of the Study:

• To develop an improved scene segmentation network by enhancing feature representation.
• To investigate and optimize the use of "skip" connections for better spatial information retention.
• To introduce novel architectural components for improved low-level and high-level feature aggregation.

Main Methods:

• Comparison of "dilation" and "skip" connection strategies for retaining spatial information from pre-trained CNNs.
• Modification of skip layer parameterization to enhance parsing performance.
• Introduction of a "dense skip" architecture for rich low-level feature retention.
• Proposal of a Convolutional Context Network (CCN) for aggregating high-level contextual information.

Main Results:

• Demonstrated significant performance improvement in "skip" networks through parameterization modification.
• The "dense skip" architecture effectively retains essential low-level information.
• The CCN successfully aggregates contexts for robust high-level parsing.
• The proposed Enhanced Fully Convolutional Network (EFCN) achieves state-of-the-art performance on ADE20K, Pascal Context, SUN-RGBD, and Pascal VOC 2012 datasets.

Conclusions:

• The EFCN architecture offers substantial improvements over standard Fully Convolutional Networks (FCNs).
• The integration of "dense skip" connections and CCN is key to EFCN's superior performance.
• EFCN provides a robust and effective solution for scene segmentation without complex add-ons.