Deep Residual Correction Network for Partial Domain Adaptation

Deep domain adaptation methods have achieved appealing performance by learning transferable representations from a well-labeled source to different but related unlabeled target domain. Most existing works assume and data share the identical label space, which is often difficult be satisfied in many real-world applications. With emergence of big data, there more practical scenario called partial adaptation, where we are always accessible large-scale while working on relative small-scale In this case, conventional assumption should relaxed, space tends subset space. Intuitively, reinforcing positive effects most relevant subclasses reducing negative impacts irrelevant vital importance address challenge. This paper proposes an efficiently-implemented Residual Correction Network (DRCN) plugging one residual block into network along with task-specific feature layer, effectively enhances explicitly weakens influence classes. Specifically, plugged block, consists several fully-connected layers, could deepen basic boost its representation capability correspondingly. Moreover, design weighted class-wise alignment loss couple two domains matching distributions shared classes between target. Comprehensive experiments partial, traditional fine-grained cross-domain visual recognition demonstrate that DRCN superior competitive deep approaches.