Transfer Sparse Subspace Analysis for Unsupervised Cross-View Scene Model Adaptation

Abstract

With the availability of an increasing amount of images from the Internet and several user-friendly crowd sourcing tools such as Amazon Mechanical Turk (AMT), many large-scale ground level image datasets with semantic annotations have been collected in the vision community, and they have fostered many efficient ways to describe the image content. For example, visual attributes have shown promising potentials in visual recognition and retrieval. However, the scarcity of labeled samples in the earth observation (EO) community (collection of labeled samples through photointerpretation or terrestrial campaigns is time consuming and expensive, and often requires expertise background) hinders the semantic understanding of remote sensing images. In this paper, we propose to transfer the semantic knowledge learned from ground view scene images to overhead view very high-resolution (VHR) remote sensing images. Specifically, a novel transfer sparse subspace analysis (TSSA) algorithm is presented for unsupervised cross-view scene modeladaptation. TSSA aims at finding a common embedding of the data across different views by simultaneously 1) minimizing the maximum mean discrepancy (MMD), 2) preserving the main statistical property, and 3) maintaining the self-expressiveness property of the data in a reproducing kernel Hilbert space (RKHS). Two variants of cross-view scene knowledge transfer have been investigated in our experiments. The first one is transfer of scene category models, and the second is transfer of scene attributes models. Experimental results demonstrate the competitive performance of our method with respect to state of the art.