Nonparametric scene parsing with deep convolutional features and dense alignment

This paper addresses two key issues which concern the performance of nonparametric scene parsing: (1) the semantic quality of image retrieval; and (2) the accuracy in label transfer. First, because nonparametric methods annotate a query image through transferring labels from retrieved images, the task of image retrieval should find a

set of "semantically similar" images to the query. Second, with the retrieval set, a good strategy should be developed to transfer semantic labels in pixel-level accuracy. In this paper, we focus on improving scene parsing accuracy in these two issues. We propose using the state-of-the-art deep convolutional features as image descriptors to improve the

semantic quality of retrieved images. In addition, we include dense alignment into the Markov Random Field inference framework to transfer labels at pixel-level accuracy. Our

experiments on the SIFT Flow dataset shows the improvement of the proposed approach over other nonparametric methods.