Semantic Correspondence using Self Supervised Learning

Abstract

Semantic correspondence is a critical task in computer vision, with applications ranging from object recognition to image matching. Despite recent advances in deep learning, the full potential of self-supervised learning (SSL) and the underlying pyramidal hierarchy of convolutional neural networks (CNNs) for semantic correspondence has yet to be fully explored. In this thesis, we present a novel approach for semantic correspondence using self-supervised learning, which effectively leverages the power of SSL and the multi-scale nature of CNNs. Our method, termed Semantic Correspondence using Self-supervised Learning (SC-SSL), employs a contrastive learning framework to learn discriminative pixel-level features by maximizing the mutual information between image pairs. By utilizing the inherent pyramidal hierarchy within CNNs, our approach captures rich, multi-scale feature embeddings that lead to improved performance on various semantic correspondence tasks. Through extensive experiments, we demonstrate that our proposed SC-SSL model achieves competitive results when compared to supervised methods.