Chapter 3 - Unsupervised Dictionary Optimization

Abstract

The core part of the learning-based local representation lies in the visual dictionary and its indexing structure, e.g., how to optimize the local feature-space quantization scheme. In Chapter 2, the design of the optimal interest-point detector was covered, while in this chapter, we introduce and discuss the way to optimize the dictionary construction-based on hierarchical metric learning unsupervised. We will show that the traditional Euclidean distance-based metric is unsuitable for k-means clustering, which would push clusterings located at dense regions and hence result in a biased distance metric. To deal with this issue, this chapter introduces density-based metric learning (DML) to refine the similarity metric in hierarchical clustering, which makes middle levels powerful and meaningful in retrieval. Based on an optimized VT model, we propose a novel idea to leverage the model hierarchy in a vertical boosting chain manner [84] to improve retrieval effectiveness. We will show that such an idea can generate a visual word distribution of the resulting dictionary to be more similar to the textual word distribution of documents. Furthermore, we also demonstrate that exploiting the VT hierarchy can improve its generativity across different databases. We propose a “VT shift” algorithm to transfer a vocabulary tree model into new databases, which efficiently maintains considerable performance without reclustering. Our proposed VT shift algorithm also enables incremental indexing of the vocabulary tree model for a changeable database in a scalable scenario. In this case our algorithm can efficiently include new data into model refinement without regenerating an entire model from the overall database.