Evaluation of Segmentation Quality via Adaptive Composition of Reference Segmentations

Bo Peng, Lei Zhang, Xuanqin Mou and Ming-Hsuan Yang

Home

Framework

Measure

Databases

Results

Publications

Motivation and Framework

Human visual system is highly adapted to extract structural information from natural scenes, and a perceptually meaningful measure should be error-sensitive to the structures in the segmentations. We propose that the evaluation of a segmentation should rely more on the local structures rather than on the global view. We assume that if a segmentation is "good", it can be composed by pieces of the ground truths.

The composite reference segmentation image should locally match the input segmentation as much as possible. For an input segmentation, a composite reference segmentation image is adaptively constructed from the ground truths in the database, and then the quality score of the segmentation is calculated by matching the input segmentation with the composite reference segmentation image (see figure below).

Proposed evaluation framework based on adaptive composition of reference segmentation images.