CS194-26 Final Project: Seam Carving & Fake Miniatures

Michael Weymouth (cs194-26-adc)

Seam Carving

For this project, we implemented the paper Seam Carving for Content-Aware Image Resizing to intelligently remove seams from an image and shrink its size. This is done by using an energy function to compute a minimum energy path through the image (left-to-right for vertical resizing and top-to-bottom for horizontal resizing), then removing all of the pixels on that path and outputting the new image. This process is repeated until the selected dimension is at the desired size in pixels.

I present below a select few success and failure examples of horizontal and vertical resizing, with the original image on the left and the carved result on the right. All photos are my own.

Successful Horizontal Resizing

The Cliffs of Moher, Ireland

The Butchart Gardens, British Columbia, Canada

Delicious Sushi

Yosemite National Park, California

Successful Vertical Resizing

Chihuly Garden and Glass, Seattle, WA

Paradise, Mount Rainier National Park, WA

Seattle, WA

St. Andrews, Scotland, UK

Unsuccessful Horizontal Resizing

The K Club, County Kildare, Ireland

Unsuccessful Vertical Resizing

Dromoland Castle, County Clare, Ireland

Victoria, British Columbia, Canada

Bells & Whistles: Visual Mode

For Bells & Whistles, I implemented a visual mode which displays the seams that the algorithm is removing as the program runs. This is extremely helpful for visualizing the seams that the algorithm chooses and evaluating different energy functions while refining the program. Below is a demonstration of this mode in action.

Most important thing I learned?

The most important thing I learned from this project was how to judge whether or not an image would work well for seam carving. An image with a clear path across in the desired direction works best, and being able to spot this quickly became an extremely valuable skill. Admittedly this is not a perfect heuristic, as the failure image The K Club above shows, but it was good enough to be helpful.

Fake Miniatures

For this project, we turned regular photographs into fake miniature scenes. This is done as follows: first, we take the square root of the saturation of the image to increase this value while preventing it from clipping. Then, we define a polygon on the image with a set of points to specify the area of perfect focus. We “expand” the polygon outwards from its center in successive steps to generate a set of blurred delta masks which represent the increase in distance from the region of focus. These blurred masks create a gentle transition between the regions of decreasing focus. At each step, we also blur the image using a Gaussian filter with increasing variance to simulate the image losing focus at farther distances from the focus area. Finally, we blend the images of increasing blur using the corresponding delta masks.

I present below a selection of several miniaturized images. The photo on the left is the original, while the image on the right is the miniaturized version. Except where otherwise specified, the images below are my own.

Manhattan Sunset #2 – New York City Skyline (Source)

The Butchart Gardens, British Columbia, Canada

Victoria, British Columbia, Canada

St. Andrews, Scotland, UK

St. Andrews, Scotland, UK

Edinburgh, Scotland, UK

Field (Source)

Bells & Whistles

As described above, I chose to implement complex DOF regions using object masks. These masks are generated by selecting a set of points to define a polygon on the original image. This method allowed for much more realistic results, so it was used for all of the results above.