This project focuses on graph-based semi-supervised learning for object recognition. We explore different algorithms and their dependency on the underlying graph structure given by different image descriptors and distance measures. We provide several unsupervised and supervised graph improvement, e.g., metric learning or dimensionality reduction. Finally, we propose a novel active learning framework RALF that make use of reinforcement learning to find a dataset-independent and good labeling.
We explore knowledge transfer and sharing in a large scale classification setting. We compare hierarchical, attribute-based, and direct transfer with one-vs-all learning.
To enable large knowledge transfer from known to unkown image classes we tab into different linguistic knowledge bases such as Wikipedia and the WWW. We examine attribute-based and direct transfer methods for transferring knowledge.