DeCAF: A Deep Convolutional Activation Feature for Generic Visual\n Recognition

Abstract

We evaluate whether features extracted from the activation of a deep\nconvolutional network trained in a fully supervised fashion on a large, fixed\nset of object recognition tasks can be re-purposed to novel generic tasks. Our\ngeneric tasks may differ significantly from the originally trained tasks and\nthere may be insufficient labeled or unlabeled data to conventionally train or\nadapt a deep architecture to the new tasks. We investigate and visualize the\nsemantic clustering of deep convolutional features with respect to a variety of\nsuch tasks, including scene recognition, domain adaptation, and fine-grained\nrecognition challenges. We compare the efficacy of relying on various network\nlevels to define a fixed feature, and report novel results that significantly\noutperform the state-of-the-art on several important vision challenges. We are\nreleasing DeCAF, an open-source implementation of these deep convolutional\nactivation features, along with all associated network parameters to enable\nvision researchers to be able to conduct experimentation with deep\nrepresentations across a range of visual concept learning paradigms.\n