“Capturing the moment”talk by Michael CohenPosted: 2009/10/31
Cornell Computer Science department has a weekly colloquium series every Thursday. I look forward to attending these talks as colloquium series held at Information Science department is not as technical. I found myself dozing several times after 15 minutes into the talks.
Two weeks ago, microsoft researcher Michael Cohen came to Cornell to give out a talk on his recent research. As my initial field of interest in my graduate study was computer graphics, I’ve been reading SIGGRAPH proceedings earlier on in my graduate program. While reading those proceedings, Michael Cohen’s work caught my interest as a lot of his work is computational photography which I have been greatly interested in.
Computational photography is a field that has the most thin boundary between computer graphics and computer vision. The goal of this field is to computationally produce an image that doesn’t exist among captured images. Instead of describing this topic in totally technical terms, just as how I described it, Michael Cohen summarized it philosophically in his talk’s title; “capturing the moment”.
Photograph is an objective visual representation of a point in time and space. Comparatively, moment is a visual representation of the subjective reality at some specific time and place. Photograph is not good at capturing this subjective moment. For example, when taking a group shot, it is very difficult to take a single shot where everybody has their eyes opened. Because photography is imperfect, computational photography proposes several solution to perfect images for capturing the moment that the photographer wants to take.
Using Groupshot, software that is based on Cohen’s 2004 SIGGRAPH publication (Interactive Digital Photomontage), users can stitch together part of photos that people like. By taking pieces of people smiling from different photos and switching them together, users can create a group shot that everyone is satisfied with.
A moment that a person wants to record might be difficult to capture not only because of the human factors but also due to limitations of camera sensors. One example is taking a night shot. When flash is on, the camera sensors capture the image at the grain level that your eyes see but not with the right color. When flash is not on, camera captures an image with the color tone that your eyes see but not with the contrast. Cohen’s solution to this problem is to use joint bilateral filter to take the color information from ‘no-flash’ image and the detail from ‘flash’ image and combine them. When I originally saw this paper from SIGGRAPH 2005 proceeding, I thought that it was a brilliant idea. When I heard him describing it again during the colloquium, I still thought that it was a brilliant idea~!
His more recent work extends what I just described in high depth range images, giga pixel images, videos, google street view images and many more. I can probably write another whole page summarizing his work that I have been following. Maybe if I run out of contents to talk about as part of this attempt to write 365 posts, I might consider writing more. Today, I think I met my requirements for my resolution. ^_^
When the talk was over, one of the audience asked him a very interesting question.
“Given all your techniques in computational photography, can you prove that a photo is real?” My guess is that he might have been involved in crime investigation relating photo forgery. Michael Cohen answered, “That is a very good question. The answer to that question is that using current computer graphics techniques, we can tell if an image has been altered. However, we cannot prove that an image is authentic.”