A Novel Upsampling and Context Convolution for Image Semantic Segmentation

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Semantic segmentation is crucial for robot vision and autonomous driving, providing object details like boundaries and locations.
• Current encoder-decoder deep convolutional neural networks for semantic segmentation often lose spatial information and lack scene context.
• This loss of detail hinders precise object boundary delineation and overall segmentation accuracy.

Purpose of the Study:

• To propose a novel semantic segmentation method that effectively preserves spatial information.
• To introduce a technique that enhances the understanding of scene context for more accurate predictions.
• To improve the delineation of object boundaries in images.

Main Methods:

• A novel dense upsampling convolution method utilizing a guided filter to retain spatial details.
• A new local context convolution method designed for dense coverage of objects, including large-scale ones.
• Integration of these methods into an encoder-decoder framework for semantic segmentation.

Main Results:

• The proposed method significantly preserves spatial information lost in traditional encoder-decoder architectures.
• New state-of-the-art pixel accuracy records were achieved on benchmark datasets: 82.86% on ADE20K and 81.62% on Pascal-Context.
• Qualitative results demonstrate superior object boundary delineation compared to existing methods.

Conclusions:

• The novel dense upsampling and local context convolution methods offer substantial improvements in semantic segmentation.
• The approach effectively addresses spatial information loss and enhances scene context comprehension.
• This work presents a promising advancement for computer vision tasks requiring precise image understanding.