Rank-Based Unsupervised Similarity Learning: Framework and Applications
Resumo
Multimedia data collections have grown exponentially due to advances in acquisition and sharing technologies, creating a pressing need for robust similarity learning methods that do not depend on costly annotations. Deep features extracted by convolutional and transformer networks provide powerful embeddings, but often lie on complex manifolds that simple pairwise comparisons fail to capture. Unsupervised similarity learning addresses this gap by post-processing these embeddings with rank-based strategies that leverage the ordering of neighbors and manifold geometry through different approaches, e.g., graph and hypergraph constructions. This paper discusses foundational concepts in contextual similarity and presents a comprehensive overview of rank-based methods for unsupervised distance and similarity learning. It discusses the Unsupervised Distance Learning Framework (UDLF), along with its Python wrapper, pyUDLF, and the web-based interface UDLFWeb, which automates and simplifies the processes of experimentation and result visualization. We also show how these tools enhance the effectiveness of image retrieval, classification, and clustering, and outline future directions for expanding rank-based methods to additional modalities and scaling them to large datasets. The complete source code and documentation for UDLF and its associated tools are publicly available at: udlf.lucasvalem.com.
