FreeStyle: Toward Style-Inclusive Sketch-Based Person Retrieval.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Sketch-based Person Retrieval (SBPR) identifies individuals across different camera views using sketches.
• Diverse artist styles in sketches challenge SBPR model stability and generalization.
• Existing methods often degrade structural sketch features, leading to poor performance on new styles.

Purpose of the Study:

• To introduce FreeStyle, a novel style-inclusive framework for robust SBPR.
• To enable dynamic adaptation to both seen and unseen sketch style variations.
• To improve the generalizability and stability of SBPR models.

Main Methods:

• FreeStyle utilizes the CLIP architecture with explicit style relation modeling.
• Diverse Style Semantic Unification enhances semantic-level style consistency using attribute constraints.
• Diverse Style Feature Squeezing improves feature representation by concentrating intra-identity and separating inter-identity spaces.
• Identity-centric cross-modal prototype alignment addresses sketch-photo feature discrepancies.

Main Results:

• FreeStyle demonstrates stable performance across diverse, seen sketch styles.
• The framework shows strong generalization capabilities on unseen sketch styles.
• Experiments validate the effectiveness of the proposed style consistency enhancement and cross-modal alignment.

Conclusions:

• FreeStyle offers a significant advancement in sketch-based person retrieval by effectively handling style variations.
• The proposed methods improve model robustness and generalization, crucial for real-world surveillance applications.
• This work paves the way for more reliable person identification systems using diverse sketch inputs.