Cross-Paced Representation Learning With Partial Curricula for Sketch-Based Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Sketch-based image retrieval (SBIR) commonly uses low- and mid-level descriptors.
• Learning coupled feature representations enhances cross-domain SBIR but faces optimization challenges.
• Existing methods struggle with non-convex problems and lack handling of prior knowledge.

Purpose of the Study:

• To develop a robust cross-domain representation learning framework for SBIR.
• To address optimization difficulties in current cross-domain methods.
• To introduce a novel approach inspired by self-paced learning (SPL).

Main Methods:

• Introduced the cross-paced partial curriculum learning (CPPCL) framework.
• CPPCL leverages self-paced learning (SPL) principles for improved convergence.
• The method jointly handles dual data sources and modality-specific prior information.

Main Results:

• CPPCL demonstrates superior performance over competing SBIR methods.
• The framework effectively embeds robust coupled representations for SBIR.
• Evaluated on four public datasets: CUFS, Flickr15K, QueenMary SBIR, and TU-Berlin Extension.

Conclusions:

• The proposed CPPCL framework offers a robust solution for cross-domain representation learning in SBIR.
• CPPCL overcomes limitations of existing SPL methods by incorporating dual-source data and prior knowledge.
• This approach significantly advances the state-of-the-art in sketch-based image retrieval.