Recent developments in deep domain adaptation have allowed knowledge transfer from a labeled source domain to an unlabeled target domain at the level of intermediate features and input pixels. We propose that advantages may be derived by jointly investigating the two, in the form of different level insights that lead to a novel design and complementary properties that result in better performance. At the higher level in feature space, inspired from semi-supervised learning, we propose a classification-aware domain adversarial neural network that brings target examples into more classifiable regions of source domain. To further boost the performance, we posit that computer vision insights are more amenable to injection at the lower level pixel space. Specifically, a general appearance flow framework is applied to transform the 3D geometry while strictly preserving identity. An attribute-conditioned CycleGAN is proposed to translate a single source into multiple target sources, differing from the low-level properties such as lighting. We validate the proposed framework on the standard UDA benchmark, as well as a challenging car recognition benchmark Compcars, in which the labeled web images and unlabeled surveillance images are extremely dissimilar. The superior performance demonstrate that our framework successfully handles both the explicitly specified nameable factors of variation in pixel space and the implicit unspecified factors in feature space.
