Feature representation based on intrinsic structure discovery in high dimensional space

In this paper, an image is regarded as a collection of image patches that can be referred to as points with certain intrinsic structures/patterns in high-dimensional space. These structures contain vital information of image features and thus provide a novel method for image feature representation. To discover these intrinsic structures, we first propose neighborhood linear embedding (NLE), an unsupervised learning algorithm, to discover neighborhood relationship and global distribution of input data simultaneously. Secondly, NLE is extended to discover the clustering structure of data by incorporating with a Euclidean distance histogram and a series of band pass filters. Finally, by combining with a dimensionality reduction technique, the discovered intrinsic structures are visualized and manipulated in low-dimensional space in the format known as embeddings. The proposed NLE allows the discovery process to adapt to the characteristics of input data. In addition, it is revealed that an image feature composed of image patches can be tracked by tracking the contour containing embeddings of the corresponding image patches