Learning Compact Visual Attributes for Large-Scale Image Classification

International audienceAttributes based image classification has received a lot of attention recently, as an interesting tool to share knowledge across different categories or to produce compact signature of images. However, when high classification performance is expected, state-of-the-art results are typically obtained by combining Fisher Vectors (FV) and Spatial Pyramid Matching (SPM), leading to image signatures with dimensionality up to 262,144 [1]. This is a hindrance to large-scale image classification tasks, for which the attribute based approaches would be more efficient. This paper proposes a new compact way to represent images, based on attributes, which allows to obtain image signatures that are typically 103 times smaller than the FV+SPM combination without significant loss of performance. The main idea lies in the definition of intermediate level representation built by learning both image and region level visual attributes. Experiments on three challenging image databases (PASCAL VOC 2007, CalTech256 and SUN-397) validate our method