Learning generic and robust feature representations with data from multipledomains for the same problem is of great value, especially for the problemsthat have multiple datasets but none of them are large enough to provideabundant data variations. In this work, we present a pipeline for learning deepfeature representations from multiple domains with Convolutional NeuralNetworks (CNNs). When training a CNN with data from all the domains, someneurons learn representations shared across several domains, while some othersare effective only for a specific one. Based on this important observation, wepropose a Domain Guided Dropout algorithm to improve the feature learningprocedure. Experiments show the effectiveness of our pipeline and the proposedalgorithm. Our methods on the person re-identification problem outperformstate-of-the-art methods on multiple datasets by large margins.
展开▼
