Saliency-GD: A TF-IDF Analogy for Landmark Image Mining

In this paper we address the problem of unsupervised landmark mining, which is to automatically discover frequently appearing landmarks from an unstructured image dataset. Landmark mining often suffers from false matches resulted from cluttered backgrounds and foregrounds, inter-class similarities, and so on. Analogous to TF-IDF in image retrieval, we propose the Saliency-GD weighting scheme of visual words, which can be easily integrated into state-of-the-art local-feature-based visual instance mining frameworks. Saliency detection provides feature weighting in image space from the attention perspective, and in feature space, the knowledge of geographic density (GD) transferred from a separate training dataset gives a multimodal selection of meaningful visual words. Experiments on public landmark datasets show that Saliency-GD weighting scheme greatly improves the landmark mining performance with increasing discrimination power of visual features.