Hybrid transformer-graph neural network feature matching based methodology for robust template object localization

Abstract

This paper proposes a novel framework that combines transformer-based and graph neural network-based feature matching techniques for accurate template object localization. The proposed pipeline consists of four main components: a hybrid features matching module, a non-linear geometric transformation module, and a bounding box refinement module. By integrating the strengths of both global contextual understanding from transformers and structural relationship modeling from graph neural networks, the method achieves improved robustness and precision in detecting and localizing objects under challenging conditions.