Separability Criteria for the Evaluation of Boundary Detection Benchmarks

Area of Science:

• Computer Vision
• Image Analysis
• Machine Learning

Background:

• Numerous benchmarks exist for evaluating boundary detection algorithms, typically comparing automated results to human-labeled data.
• These benchmarks comprise image datasets and comparison measures, but optimal combinations for evaluating algorithm performance remain unclear.

Purpose of the Study:

• To introduce four novel criteria for evaluating the performance of comparison measures within boundary detection benchmarks.
• To provide a framework for assessing the suitability of benchmark-dataset combinations for evaluating boundary detection methods.

Main Methods:

• Development of four criteria designed to emulate human perception of boundary images and scene recognition.
• Application of these criteria to evaluate the effectiveness of existing comparison measures on specific datasets.

Main Results:

• The proposed criteria offer a sensible way to evaluate the performance of comparison measures on boundary detection datasets.
• These criteria can quantify the efficacy of benchmarks in assessing both edge-based and segmentation-based boundary detection algorithms.

Conclusions:

• The introduced criteria enhance the evaluation of boundary detection benchmarks.
• This work provides a pathway to more accurately quantify the performance of boundary detection algorithms through improved benchmark assessment.