Support Vector Guided Locally Aggregated Descriptors for Feature Encoding

Feature encoding, as a crucial component of image categorization, has been widely investigated in recent works. Among all kinds of encoding strategies, the vector of locally aggregated descriptors (VLAD) and its extensions have gained a lot of attentions and provide a brand-new framework, aggregating the deviations between a set of descriptors and their average distribution. However, the VLAD-like frameworks only utilize statistics to represent the deviations, which may not achieve the optimal discrimination for classification tasks. In this paper, we propose the support vector guided locally aggregated descriptors (SVGLAD) to leverage the residuals of support vectors between descriptors and clusters of the dataset, which can preserve more geometric information. We validate the effectiveness of our methods on two image datasets (15-Scenes and PASCAL VOC 2007). The experimental results demonstrate that the proposed method outperforms the existing coding algorithms.