Cross-Dataset Learning of Visual Concepts

Christian Hentschel, Harald Sack, Nadine Steinmetz

Abstract: Visual content classification has become a keystone when opening up digital image archives to semantic search. Content-based explicit metadata often is only sparsely available and automated analysis of the depicted content therefore provides an important source of additional information. While visual content classification has proven beneficial, a major concern, however, is the dependency on large scale training data required to train robust classifiers. In this paper, we analyze the use of cross-dataset training samples to increase the classification performance. We investigate the performance of standardized manually annotated training sets as well automatically mined datasets from potentially unreliable web resources such as Flickr and Google Images. Next to brute force learning using this potentially noisy ground truth data we apply semantic post processing for data cleansing and topic disambiguation. We evaluate our results on standardized datasets by comparing our classification performance with proper ground truth-based classification results.