Y.Guo et aL.Computers Graphics 38(2014)174-182 181 Input Lightroom Ours Fig.12.Upright adjustment for a photo of the Taj Mahal (left).an indoor image of a temple (middle).and a church photo(right).The 2nd row shows the results by Adobe Lightroom Upright.and the 3rd row shows our results.The results of Lightroom show apparent visual artifacts (see the red ellipses in the 1st and 3rd results).(For interpretation of the references to color in this figure caption.the reader is referred to the web version of this article.) changing of the environment is not supported.Second,and more importantly,to meet the different goals,only partial and inaccu- rate recovery of the standard and non-standard cuboid structures is required by our approach.We do not need to accurately reconstruct the geometry of the cuboid structure.Only five edges on the cuboid dominated structure for an input image,including a so-called vertical edge and two pairs of parallel edges (not necessarily of identical length)are enough.This is validated by most of our experimental results.We believe it is challenging to automatically detect the cuboid structures and reconstruct the accurate geometry for most our testing images,such as most interest objects in Figs.6 and 7 and the indoor images in Fig.8,by the algorithm in [6.Furthermore,our approach is competent for viewpoint manipulations for the image with a non- standard cuboid dominated structure like the input photo shown in Fig.9. 5.2.Upright adjustment Man-made structures often appear to be slanted in photos taken by casual photographers.An example is shown in the 1st column of Fig.12.This is partly due to the improper position where the camera is placed at.Human visual system however Fig.13.Images under new viewpoints are interpolated from the original photo and always expects tall man-made structures to be straight-up.In [3]. four new images under key viewpoints highlighted with red rectangles.(For the slanted structures are dealt with by using an improved interpretation of the references to color in this figure caption,the reader is referred to the web version of this article. homography model.Our algorithm can also be used to straighten up the slanted cuboid structures in images.The idea is to re- generate the image by modifying the viewpoint of the input image properly.The new viewpoint is computed automatically by letting show apparent visual artifacts in the results of Taj Mahal and the projected structures to be vertical. Church photos (see the red ellipses).In [3].the authors assume A software implementation of [3]is Adobe Lightroom Upright. that depth variations of the scene relative to its distance from the We thus compare our results to those produced by Lightroom. camera are small.The reason for the artifacts may be that such Fig.12 shows the results produced by Lightroom (2nd row)and assumption does not hold exactly for the two images.Transforming those by our algorithm (3rd row).Our results are generally the input image with a homography is not always sufficient since it comparable to those by Lightroom upright.The results of Lightroom is oblivious to the depth variations of the latent scenes.changing of the environment is not supported. Second, and more importantly, to meet the different goals, only partial and inaccu￾rate recovery of the standard and non-standard cuboid structures is required by our approach. We do not need to accurately reconstruct the geometry of the cuboid structure. Only five edges on the cuboid dominated structure for an input image, including a so-called vertical edge and two pairs of parallel edges (not necessarily of identical length) are enough. This is validated by most of our experimental results. We believe it is challenging to automatically detect the cuboid structures and reconstruct the accurate geometry for most our testing images, such as most interest objects in Figs. 6 and 7 and the indoor images in Fig. 8, by the algorithm in [6]. Furthermore, our approach is competent for viewpoint manipulations for the image with a non￾standard cuboid dominated structure like the input photo shown in Fig. 9. 5.2. Upright adjustment Man-made structures often appear to be slanted in photos taken by casual photographers. An example is shown in the 1st column of Fig. 12. This is partly due to the improper position where the camera is placed at. Human visual system however always expects tall man-made structures to be straight-up. In [3], the slanted structures are dealt with by using an improved homography model. Our algorithm can also be used to straighten up the slanted cuboid structures in images. The idea is to re￾generate the image by modifying the viewpoint of the input image properly. The new viewpoint is computed automatically by letting the projected structures to be vertical. A software implementation of [3] is Adobe Lightroom Upright. We thus compare our results to those produced by Lightroom. Fig. 12 shows the results produced by Lightroom (2nd row) and those by our algorithm (3rd row). Our results are generally comparable to those by Lightroom upright. The results of Lightroom show apparent visual artifacts in the results of Taj Mahal and Church photos (see the red ellipses). In [3], the authors assume that depth variations of the scene relative to its distance from the camera are small. The reason for the artifacts may be that such assumption does not hold exactly for the two images. Transforming the input image with a homography is not always sufficient since it is oblivious to the depth variations of the latent scenes. Input Lightroom Ours Fig. 12. Upright adjustment for a photo of the Taj Mahal (left), an indoor image of a temple (middle), and a church photo (right). The 2nd row shows the results by Adobe Lightroom Upright, and the 3rd row shows our results. The results of Lightroom show apparent visual artifacts (see the red ellipses in the 1st and 3rd results). (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this article.) Fig. 13. Images under new viewpoints are interpolated from the original photo and four new images under key viewpoints highlighted with red rectangles. (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this article.) Y. Guo et al. / Computers & Graphics 38 (2014) 174–182 181