Spatio-temporal video segmentation with shape growth or shrinkage constraint.

Area of Science:

• Computer Vision
• Image Processing
• Computational Imaging

Background:

• Accurate segmentation of dynamic shapes in image sequences is challenging, especially with noise.
• Existing methods often struggle to enforce constraints on shape evolution over time.

Purpose of the Study:

• To introduce a new method for joint segmentation of monotonously growing or shrinking shapes in time-series image data.
• To leverage spatio-temporal graph optimization for robust shape tracking.

Main Methods:

• Formulating image time-series segmentation as a graph optimization problem.
• Introducing monodirectional infinite links in a spatio-temporal pixel graph to enforce shape growth/shrinkage constraints.
• Utilizing graph cuts for computing the globally optimal segmentation solution.
• Incorporating inter-sequence inclusion constraints for dependent structures in 3D medical scans.

Main Results:

• Successfully segmented melting sea ice floes from 2D satellite images.
• Accurately segmented growing burned areas from time-series 2D satellite imagery.
• Effectively segmented a growing brain tumor from 3D medical scan sequences.

Conclusions:

• The proposed graph cut method provides an effective solution for segmenting evolving shapes in noisy image sequences.
• The method demonstrates versatility across diverse applications, including remote sensing and medical imaging.
• The incorporation of specific constraints, like shape monotonicity and inter-sequence inclusion, enhances segmentation accuracy.