Context Matters: Effective Context Mining for Object Detection and Classification

This thesis develops innovative context mining methods to address the ambiguity in visual data for object-level visual understanding and the lack of inherent structure in bioinformatic data for cell classification. A context-aware knowledge graph for object-level visual understanding is developed to effectively resolve ambiguities in identifying objects of interest. A structural gene tree method constructed through unsupervised min-max optimization is developed to overcome the challenge of lacking natural structure and domain knowledge in scRNA-seq cell classification. These developments represent original contributions to visual understanding research using visual data and classification research using bioinformatic data, with both methodological and practical significance.