Supervoxel Segmentation with Voxel-Related Gaussian Mixture Model

Area of Science:

• Computer Vision
• Image Processing
• Video Analysis

Background:

• Supervoxel segmentation partitions video frames into atomic segments.
• Existing methods struggle with temporal consistency, especially for new and moving objects.

Purpose of the Study:

• To propose a novel supervoxel segmentation scheme addressing new and moving objects.
• To enhance temporal consistency and robustness in video segmentation.

Main Methods:

• Segmentation performed on consecutive frames, yielding two superpixel segmentations per internal frame.
• A voxel-related Gaussian mixture model (GMM) representing supervoxels with two shared-color Gaussian distributions.
• Algorithm designed for lower complexity relative to frame size compared to traditional GMM.

Main Results:

• The proposed scheme offers coarse-grained parallel ability for segmentation.
• Subsequent algorithms can leverage dual segmentations for improved robustness.
• Experimental results demonstrate superior accuracy compared to state-of-the-art methods.

Conclusions:

• The novel supervoxel segmentation scheme effectively handles new and moving objects.
• The voxel-related GMM improves temporal consistency and computational efficiency.
• This method advances video segmentation accuracy and robustness.