A Novel Semi-Supervised Learning Approach for Referring Expression Comprehension
Lijie Geng, Huiyi Liu, Tengfei Wan · 2024
Referring Expression Comprehension (REC) aims to interpret natural language expressions and associate them with corresponding objects in image raditional REC methods often rely on a two-stage process, which can be inefficient and less accurat n this paper, we introduce RecGLIP, a semi-supervised learning model that employs a teacher-student framework for single-stage REC. Our approach leverages mutually optimized strategies, including pretraining on image-text pairs and a deep fusion strategy, to enhance mode l efficiency and accurac he proposed model is tested on benchmark datasets such as RefCOCO, RefCOCO+, and RefCOCOg, demonstrating superior generalization capabilitie otably, RecGLIP achieves significant performance improvements, showing over a 10% increase in accuracy compared to existing semi-supervised REC methods, particularly in scenarios with limited labeled dat his advancement showcases the potential of semi-supervised learning in overcoming the challenges of extensive instance-level annotation in REC, paving the way for more effective and practical applications.