Co-transduction for shape retrieval.

Area of Science:

• Computer Science
• Machine Learning
• Image Analysis

Background:

• Shape retrieval systems rely heavily on accurate similarity measures.
• Designing ideal metrics that handle large intraclass variations in object shapes is challenging.
• Complementary information from different measures (e.g., contours, skeletons) can improve retrieval robustness.

Purpose of the Study:

• To develop a robust shape retrieval algorithm by fusing diverse similarity measures.
• To leverage a semisupervised learning framework for enhanced retrieval performance.
• To introduce a generalizable method applicable beyond shape retrieval.

Main Methods:

• Propose "co-transduction," a semisupervised algorithm inspired by co-training.
• Iteratively use one similarity measure to retrieve shapes, then re-rank them using another measure, and vice versa.
• Introduce "tri-transduction" to fuse multiple similarity measures.

Main Results:

• Co-transduction achieved 97.72% bull's-eye measure on the MPEG-7 dataset, surpassing state-of-the-art.
• Tri-transduction further improved performance to 99.06% on the MPEG-7 dataset.
• The algorithm demonstrates significant improvements in robust shape retrieval.

Conclusions:

• The proposed co-transduction algorithm effectively fuses multiple similarity measures for robust shape retrieval.
• The method is general and can be applied to various ranking and retrieval tasks.
• Semisupervised learning provides a powerful framework for enhancing retrieval system accuracy.