A Testbed for Cross-Dataset Analysis

Despite the increasing interest towards domain adaptation and transfer learning techniques to generalize over image collections and overcome their biases, the visual community misses a large scale testbed for cross-dataset analysis. In this paper we discuss the challenges faced when aligning twelve existing image databases in a unique corpus, and we propose two cross-dataset setups that introduce new interesting research questions. Moreover, we report on a first set of experimental domain adaptation tests showing the effectiveness of iterative self-labeling for large scale problems.