Sparse Embedding-based Domain Adaptation for Object Recognition

Domain adaptation algorithms aim at handling the shift between source and target domains. A classifier is trained on images from the source domain; and the classifier recognizes objects in images from the target domain. In this paper, we present a joint subspace and dictionary learning framework for domain adaptation. Our approach simultaneously exploits the low-dimensional structures in two domains and the sparsity of features in the projected subspace. Specifically, we first learn domain-specific subspaces from the source and target domains respectively that can decrease the mismatch between source and target domains. Then we project features from each domain onto their domainspecific subspaces. From the projected features, a common domain-invariant dictionary for both domains is learned. Our approach handles domain shift caused by different classes of features; e.g., SURF and SIFT. In addition, the features can have different dimensions. Our framework applies to both cross-domain adaptation (cross-DA) and multiple source domain adaptation (multi-DA). Our experimental results on the benchmark dataset show that our algorithm outperforms the state of the art. Suppose that we have P domains, and the first P − 1 domains are source domains while the last domain is the target domain. Let Xπ ∈ Rdπ×Nπ denote the feature representation of Nπ samples in the π-th domain, where each column in Xπ denotes a sample and dπ is the feature dimension in the π-th domain. Our goal is to jointly learn domain-specific subspaces Wπ ∈ Rm×dπ for each domain and a common domain-invariant dictionary D ∈ Rm×J . The objective function is formulated as follows: