Human Part Segmentation in Depth Images with Annotated Part Positions.

Area of Science:

• Computer Vision
• Machine Learning
• Human Pose Estimation

Background:

• Accurate per-pixel labeling of human images is crucial for training pose estimation and segmentation algorithms.
• Existing image segmentation methods often struggle with occlusions, particularly from body parts like arms.

Purpose of the Study:

• To develop an effective method for segmenting human body parts in depth images.
• To improve the accuracy of per-pixel labeling for large datasets used in computer vision tasks.

Main Methods:

• A layered graph representation is introduced to model occlusions in depth images.
• Annotated body part positions are used as seeds for an interactive segmentation algorithm.
• The method utilizes a grid graph with distinct layers of nodes to handle occlusions.

Main Results:

• Achieved a mean per-class accuracy of 93.55% on the first dataset, outperforming existing methods.
• Obtained a per-class accuracy of 90.60% on the second dataset.
• Demonstrated superior performance compared to random forest with graph cuts and Markov random fields.

Conclusions:

• The proposed layered graph segmentation method effectively handles human body part segmentation in depth images.
• The approach significantly enhances accuracy for per-pixel labeling tasks.
• Future research can explore advanced graph layer construction for improved occlusion modeling.