The results in this supplementary material are divided into four sections.

In Big Panoramas, we show comparisons between the visual results and offset fields produced by the [Agarwala 2007] quad-tree based technique and our approach. Each thumbnail links to full-resolution versions of the offset images and half-resolution versions of the composited color images. The offset images are created by centering a zero offset at 128 and multiplying the offsets by 10 to better fill the gamut. In the interest of space, the images are JPG-compressed and the composited color images are only half-resolution. The names of each data set corresponds to the names in the Table 1 of our paper. Note that in order to save download space, the full-resolution images for the previous algorithm [Agarwala 2007] are not included in the download ZIP file. Instead, they link to Dr. Agarwala's Web site, so you will have to be connected to the Internet to view (or download) them.

In the second section, Effect of grid spacing, we compare our results, at a variety of grid sizes, to those of the [Agarwala 2007] quad-tree based technique for the 9.7MB St.Emilion dataset.  Note how even through the offset fields are visually quite different, it is almost impossible to tell the difference between the final blended images.

The third section on Outliers shows how using a power law model for the compensation, adding gradient and hue consistency outlier (weighting) terms, and adding 3 levels of Laplacian pyramid blending, all help to reduce visible seam artifacts.

The final section on Multiplicative gain vs. additive offset shows the difference between using our proposed multiplicative gain (log intensity) compensation and using traditional additive offset Poisson blending. As you can see, the additive blending results in a noticeable contrast difference across the seam boundary.

Big panoramas [Agarwala 2007]

Unblended composite

Quadtree-based composite and additive offset

Multi-spline-based composite and additive offset

Unblended composite

Quadtree-based composite and additive offset

Multi-spline-based composite and additive offset

Unblended composite

Quadtree-based composite and additive offset

Multi-spline-based composite and additive offset

Unblended composite

Quadtree-based composite and additive offset

Multi-spline-based composite and additive offset

Offset fields and blends as a function of grid size

Unblended composite

Additive offset

To see the various offset corresponding to the full Poisson blend, the quadtree approximation, and our multi-spline approximations, mouse over (but don't click!) the words below the image.

If nothing happens, check to see if one of these warnings



or



are present at the top or bottom of your browser, and if so, enable scripts for this page.

Compare: full Poisson, quadtree, S = 8, 16, 32, 64, 128.

Blended results

To see the blended results corresponding to the full Poisson blend, the quadtree approximation, and our multi-spline approximations, mouse over (but don't click!) the words below the image.  Notice that while the differences between the offset images are quite visible (aided by the 10x range magnification), the differences between the blended images are imperceptible.

Compare: full Poisson, quadtree, S = 8, 16, 32, 64, 128.

To see the full resolution offset and blended images, please look in the following directory: emilion.

Panoramic data set with outliers along the seams

Unblended composite

Final result using power(0.5) domain and both gradient and hue outlier tests.

Cropped regions

To see the benefits of various algorithm parameters mouse over their descriptions below to blink between the unblended source and a result. Notice the progressive improvement going from left to right mouse-over. The additive model doesn't fit this data set as well as the power(0.5) model, especially in the second crop. Next the gradient outlier detection improves the result by removing constraints that don't fit our model. The hue test subtly improves the result a bit more. Finally the remaining discrepencies are removed with a narrow 3 banded laplacian blend.

Compare: unblended, additive, multiplicative, power(0.5), +gradient weight, +hue weight, +3 band laplace.

Compare: unblended, additive, multiplicative, power(0.5), +gradient weight, +hue weight, +3 band laplace.

Offsets on the Additive, Multplicative, and Power Models

The following four images show the difference between using an additive offset, multiplicative gain, and an offset in the power(0.5) domain. The additive offset does not compensate for the higher contrast visible in the brighter left most image, so the colors to the right of the seam appear more muddied, even though their overall luminance is correct. The multiplicative and power models do a better job preserving the contrast in the result.

Unblended	Additive offset	Multiplicative gain	Power(0.5) offset