arXiv (Cornell University) · 2018 · 160 citations · 0 references
ComputingData ScienceFeature EngineeringKnowledge TransferMachine LearningMulti-task LearningRepresentation LearningBenchmark DatasetsFeature AlignmentFeature Representation LearningLearned Feature EmbeddingsLearned Feature RepresentationsMulti-task Feature LearningMulti-task Neural NetworksPerson Re-identification
Most existing person re-identification (Re-ID) approaches follow a supervised\nlearning framework, in which a large number of labelled matching pairs are\nrequired for training. Such a setting severely limits their scalability in\nreal-world applications where no labelled samples are available during the\ntraining phase. To overcome this limitation, we develop a novel unsupervised\nMulti-task Mid-level Feature Alignment (MMFA) network for the unsupervised\ncross-dataset person re-identification task. Under the assumption that the\nsource and target datasets share the same set of mid-level semantic attributes,\nour proposed model can be jointly optimised under the person's identity\nclassification and the attribute learning task with a cross-dataset mid-level\nfeature alignment regularisation term. In this way, the learned feature\nrepresentation can be better generalised from one dataset to another which\nfurther improve the person re-identification accuracy. Experimental results on\nfour benchmark datasets demonstrate that our proposed method outperforms the\nstate-of-the-art baselines.\n