Repairing bad co-segmentation using its quality evaluation and segment propagation.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Co-segmentation aims to identify common objects across multiple images.
• Existing methods often produce imperfect segmentations requiring improvement.

Purpose of the Study:

• To enhance co-segmentation performance by repairing inaccurate segments.
• To introduce a novel approach for evaluating and propagating segment quality.

Main Methods:

• Developed a co-segmentation quality evaluation metric based on foreground consistency and segment completeness.
• Implemented global and local region propagation techniques to transfer quality information.
• Integrated the proposed method with four state-of-the-art co-segmentation algorithms.

Main Results:

• The quality evaluation method effectively identifies good and bad segments.
• The proposed repair strategy significantly improves co-segmentation accuracy.
• Experimental results show increased intersection-over-union scores on public datasets (ICoseg, MSRC).

Conclusions:

• The novel quality evaluation and segment repair method effectively enhances co-segmentation.
• This approach offers a significant performance boost for existing co-segmentation techniques.
• The method demonstrates robustness and effectiveness across different datasets and algorithms.