Grid-Based Retargeting with Transformation Consistency Smoothing

Abstract

Effective and Efficient retargeting are critical to improve user browsing experiences in mobile devices. One important issue in previous works lies in their semantic gap in modeling user focuses and intensions from low-level features, which results to data noise in their importance map constructions. Towards noise-tolerance learning for effective retargeting, we propose a generalized content aware framework from a supervised learning viewpoint. Our main idea is to revisit the retargeting process as working out an optimal mapping function to approximate the output (desirable pixel-wise or region-wise changes) from the training data. Therefore, we adopt a prediction error decomposition strategy to measure the effectiveness of the previous retargeting methods. In addition, taking into account the data noise in importance maps, we also propose a grid-based retargeting model, which is robust and effective to data noise in real time retargeting function learning. Finally, using different mapping functions, our framework is generalized for explaining previous works, such as seam carving [9,13] and mesh based methods [3,18]. Extensive experimental comparison to state-of-the-art works have shown promising results of the proposed framework.