Evaluation of Segmentation Quality via Adaptive Composition of Reference Segmentations

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Automatic evaluation of image segmentation is crucial but difficult due to the ill-posed nature of the problem.
• Existing methods often struggle with the variability inherent in segmentation tasks.

Purpose of the Study:

• To develop a robust framework for evaluating image segmentation quality.
• To introduce a novel method for creating a composite reference segmentation from multiple labeled inputs.
• To establish a benchmark dataset for quantitative comparison of segmentation evaluation algorithms.

Main Methods:

• A framework is proposed that adaptively composes a reference segmentation from multiple labeled segmentations.
• The composed reference locally matches input segments while maintaining structural consistency.
• Segmentation quality is measured by the distance to this composed reference.

Main Results:

• A new dataset with 200 images, each featuring 6 to 15 labeled segmentations, was created for evaluation.
• The proposed framework was validated through extensive experiments.
• A benchmark dataset was proposed for quantitative comparison against state-of-the-art methods.

Conclusions:

• The proposed framework offers a reliable method for assessing image segmentation quality.
• The new dataset and benchmark facilitate objective performance evaluation of segmentation algorithms.
• This approach addresses the challenges of automatic evaluation in image segmentation.