On scene segmentation and histograms-based curve evolution.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Traditional image segmentation often relies on bin-wise metrics like Bhattacharyya or Kullback-Leibler, which can be suboptimal for complex feature distributions.
• Analyzing dynamic scenes and activities requires robust methods for segmenting spatial regions based on temporal changes.

Purpose of the Study:

• To propose and evaluate the use of cross-bin metrics, specifically Earth Mover's Distance (EMD), for feature distribution comparison in curve evolution for segmentation.
• To apply derived flow equations for semi-supervised scene segmentation in video data, focusing on activity recognition.

Main Methods:

• Developed flow equations for minimizing functionals involving EMD by using a tractable expression for 1D distributions.
• Applied these flows to single image segmentation and video scene analysis.
• Utilized a non-parametric local representation of scene regions via histograms of normalized spatiotemporal derivatives.

Main Results:

• Demonstrated effective single image segmentation using EMD-based curve evolution.
• Achieved semi-supervised segmentation of activities in challenging surveillance scenes.
• Showcased favorable comparisons against state-of-the-art methods using parametric representations.

Conclusions:

• Cross-bin metrics like EMD offer advantages over traditional bin-wise metrics for feature distribution comparison in segmentation tasks.
• The proposed EMD-based curve evolution method provides a powerful tool for semi-supervised scene and activity segmentation in video analysis.