Evaluating CNN and hybrid approaches for landmark recognition & retrieval across diverse models
Jiaming Mei, Ruoxi Wang, Xu Li +2 more
Landmark recognition and retrieval play an important role in reality, tourism, and historical site identification. This review evaluates various CNN and hybrid models on large-scale datasets, including Google Landmark v2, ROxford5k, and RParis6k. Our research shows that strategies like multi-scale feature extraction, fine-tuning, and hybrid approaches can significantly enhance recognition accuracy. The CNN with re-ranking strategy stands out from the models we examined with a mAP of 83.42% on ROxford5k and 91.80% on RParis6k under Medium difficulty. This study identifies key factors that contribute to effective landmark recognition and provides insights for optimizing models for real-time use.