Self-Supervised Visual Representation Learning Framework for Road Sign Classification
Godson Michael D’silva, Vinayak Ashok · 2023
In the landscape of machine learning, the dominance of supervised learning in crafting specialized models for specific tasks is evident, yet its confinement in terms of broader generalization has underscored the quest for more versatile models capable of acquiring multiple skills without an insurmountable reliance on labelled data. A pivotal avenue toward this goal lies in self-supervised learning (SSL), which promises to instill foundational knowledge and a semblance of common sense into AI systems. This investigation undertakes a comprehensive exploration of diverse SSL strategies, elucidating their inherent constraints on the road sign detection and classification usecase. Through an intricate comparative analysis spotlighting Context Autoencoders (CAE) and MoCoV3 - two prominent state-of-the-art methodologies leveraging masked image modelling and contrastive learning, respectively - we unravel their distinct attributes and limitations. Furthermore, we introduce an innovative end-to-end framework meticulously tailored for self-supervised road sign classification. Our roadmap encompasses the infusion of domain-specific expertise in road signs into this framework, followed by rigorous benchmarking against state-of-the-art SSL models across diverse road sign datasets, employing a spectrum of evaluation metrics. This study emerges as a stepping stone, propelling AI closer to comprehesive intelligence by harnessing the potency of self-supervised paradigms.