Depth Density Achieves a Better Result for Semantic Segmentation with the Kinect System

Area of Science:

• Computer Vision
• Animal Science
• Machine Learning

Background:

• Image segmentation is crucial for animal phenome research.
• Deep learning, particularly convolutional neural networks, is widely used for image segmentation.
• Existing methods struggle with pixel-level accuracy in segmentation.

Purpose of the Study:

• To improve the accuracy of semantic segmentation in animal image analysis.
• To introduce and validate the concept of 'depth density' for enhanced segmentation.
• To leverage depth information from Kinect systems for better phenotyping.

Main Methods:

• Developed a novel 'depth density' function using depth images from a Kinect system.
• Integrated depth density values into the semantic segmentation process.
• Utilized Fully Convolutional Networks (FCN) for image segmentation of Simmental cattle.

Main Results:

• Depth density improved semantic segmentation accuracy across four key metrics.
• Significant improvements observed: pixel accuracy (+2.9%), mean accuracy (+0.3%), mean intersection over union (+11.4%), and frequency weight intersection over union (+5.02%).
• Depth information demonstrably enhances FCN semantic segmentation performance.

Conclusions:

• Depth density is an effective method for improving pixel-level semantic segmentation accuracy.
• Kinect-derived depth data offers a valuable new approach for animal phenome analysis.
• This technique provides a novel pathway for detailed analysis of animal phenotype information.