SIGGRAPH Day 2 (David's Blog)

I started the day with a papers session called "Image Analysis & Enhancement." The first talk ("Image Deblurring With Blurred/Noisy Image Pairs") was about how to use a blurry long exposure image with a noisy short exposure image to get a sharp, clear picture with great color reproduction. My camera has image stabilization...it is not blurry, but it does contain a lot of noise. This technique would give you great pictures even when you hands are shaking or in low light situations. Very Impressive...I want this in my next camera!

Two other papers in this session were related. The first, "Photo Clip Art", showed how to add objects (like people or cars) to a photo by using MIT's LabelMe photo database to find candidates that matched the correct size and color properties of the original photo. The idea is to mimic the clip art added to PowerPoint a presentation, but for photos.

The other paper, "Scene Completion Using Millions of Photographs", let you remove objects from a photo (like an ex-girlfriend) by searching Flickr for photos that would fill in the hole nicely.

The next paper session I went to was called "Image Slicing and Stretching." The first paper ("Soft Scissors: An Interactive Tool for Realtime High Quality Matting") showed how you could just quickly trace the outline of a cat, for instance, and all the hairs that blend into the background were selected correctly.

The next paper was called "Seam Carving for Content-Aware Image Resizing." Check out the video to see it in action. Basically, they just get rid of the least important parts of a picture until it is shrunk to the desired resolution. To increase resolution, the least important parts of the picture are duplicated. The results are surprisingly good and fast. This only works for images, not video yet. Could this be the end of black bars and stretched/compressed people on TV's?

"Image Vectorization using Optimized Gradient Meshes" is a paper that deals with converting a photo into a vector format (like what is used by Adobe Illustrator or Flash animations). I have dealt with shapes as vectors, but never photos. With Windows Vista supporting vectors in the UI (to support arbitrary dpi settings) and the rising popularity of flash, storing information in vectors will become more important.

Later, I went to the sketch "Highlights from I3D". One paper presented "Tile Trees"...an alternative to normal texture mapping. The advantage is that you don't have UV's to adjust to fit a flat texture on a surface. It also does not waste any space of a texture map (like in this texture atlas).