This supplementary material contains three parts. First, we give qualitative comparisons between the importance maps generated by our method and other baseline methods. Second, we show more results of our retargeting systems. Third, we present more information on our experimental results over the Amazon Mechanical Turk (AMT) platform. In all of our experiments, the original images are resized to half of their widths while keeping their heights unchanged.

In this section, we provide several qualitative comparisons of importance maps. These importance maps contain the maps generated by our method and 6 other methods, including MC [13], GC [2], RCC [3], importance map used in the original IF (oriIF) [4], eDN [10], and DNEF [9]. We selected 30 Images from the new proposed Semantic Retarget dataset, specifically 5 images from each of the 6 categories, i.e. single person, multiple people, single object, multiple objects, indoor scene, and outdoor scene. Comparisons are listed below.

We feed our importance map to the carrier IF [4] which is selected because of its excellent performance and fast speed. We present more comparisons between our retarget results and 6 other retarget systems. These baseline retarget systems are listed below.

**SOAT** [5]: SOAT is a method to effectively obtain image thumbnails based on cropping and warping. It models the human
perception of thumbnail by visual acuity theory, and gets the Scale and Object Aware Saliency (SOAS) first, and then it uses
SOAS to do image thumbnailing.

**ISC** [7]: Improved Seam Carving (ISC) is the improved version of the famous Seam Carving [1] approach for retargeting.
For the improvement, instead of using the dynamic programming approach of seam carving, ISC prefers to graph cuts. A
novel energy criterion is also used in ISC to improve the visual quality of the retargeted images.

**Multi-Op** [8]: Multi-Operator (Multi-Op) approach provides a hybrid method of doing image retargeting. Multiple operators,
including seam carving, cropping and scaling, are used alternatively to produce the results. An image similarity
measure, named Bi-Directional Warping, is used to find the optimal path in the retargeting space.

**Warp** [12]: Warp is a warping approach for retargeting. It first analyzes the importance of each region, and then it applies
a transformation which shrinks less important regions more than important ones. This method can work both on images and
videos.

**AAD** [6]: Axis-Aligned Deformation (AAD) is a robust image retargeting method. To avoid harmful visual distortions,
deformations in AAD are parameterized in 1-dimension. Due to this 1-dimension parameterization, AAD only needs solving
a small quadratic program, so AAD method is very efficient.

**OSS** [11]: Optimized Scale-and-Stretch (OSS) is a warping method which can retarget images into any aspect ratios without
harmful visual distortions. OSS works through iteratively computing optimal local scaling factors for each localized region
and then updating warped images to match these scaling factors as closely as possible. An efficient formulation for the
nonlinear optimization is also developed to do the warping function computations.

Results are given below.

All quantitative comparisons in the paper are carried on the Amazon Mechanical Turk (AMT). Our target image and the result by a baseline are shown in randomly order to the AMT workers, who are asked to select the better one. Each pair is compared by 3 different AMT workers, and we record the numbers of votes preferring our result than the baseline. The comparison results are shown in the form of “A(B)” which means that in the total (A+B) comparisons, our method wins A times while baseline wins B times. The larger gap between A and B means more advantage.

**Table 1. **
Comparion between our importance map and 6 baseline maps when combined with 3 different carriers.

We also collected information about the workers in our experiments on AMT. Statistical information are given below.

- S. Avidan and A. Shamir. Seam carving for content aware image resizing. In TOG, 2007.
- M.-M. Cheng, N. J. Mitra, X. Huang, P. H. Torr, and S.-M. Hu. Global contrast based salient region detection. TPAMI, 2015.
- M.-M. Cheng, J. Warrell, L. Wen-Yan, S. Zheng, V. Vineet, and N. Crook. Efficient salient region detection with soft image abstraction. 2013.
- Y. Ding, J. Xiao, and J. Yu. Importance filtering for image retargeting. 2011. CVPR.
- S. Jin and L. haibin. Scale and object aware image thumbnailing. IJCV, 2013.
- D. Panozzo, O. Weber, and O. Sorkine. Robust image retargeting via axis-aligned deformation. In EUROGRAPHICS, 2012.
- M. Rubinstein, A. Shamir, and S. Avidan. Improved seam carving for video retargeting. In TOG, 2008.
- M. Rubinstein, A. Shamir, and S. Avidan. Multi-operator media retargeting. TOG, 2009.
- C. Shen, M. Song, and Q. Zhao. Learning high-level concepts by training a deep network on eye fixations. DLUFL Workshop, in conjunction with NIPS, 2012.
- E. Vig, M. Dorr, and D. Cox. Large-scale optimization of hierarchical features for saliency prediction in natural images. In CVPR, 2014.
- Y.-S. Wang, C.-L. Tai, O. Sorkine, and T.-Y. Lee. Optimized scale-and-stretch for image resizing. TOG, 2008.
- L. Wolf, M. Guttmann, and D. Cohen-Or. Non-homogeneous content-driven video-retargeting. 2007. ICCV.
- R. Zhao, W. Ouyang, H. Li, and X. Wang. Saliency detection by multi-context deep learning. In CVPR, 2015.