Is it Hollywood or is it for real?
This is one of the question that I have wondered ever since watching the first Terminator movie where Arnold, aka Mr Governor, identified his victim by analyzing video images. I am glad to report that, with no casualty count, there is indeed similar technology at work at PARC.
I have covered some of the details on this entry Video Analysis and Responsive Mirror part II. Now, I've got Maurice Chu, who worked on and modeled this technology, to share a clip.
Face tracking on a video feed
Maurice's explanation: The clip demonstrates the processing of video to track the face and its parts. In particular, the six face parts are the left and right eyes, the left and right eyebrows, nose, and mouth. The output of the face tracker is a cloud of points representing the location of the six face parts in image coordinates. These can be used to determine the 3D orientation of the face relative to the camera. Some of the challenges that the algorithm overcomes is that it can handle when people put on glasses, eye blinking, and other deformable movements like the lips. The algorithm runs in real-time, currently at about 10 fps, is people-generic, and requires no initial calibration.
===
P@P
Showing posts with label video analysis. Show all posts
Showing posts with label video analysis. Show all posts
Thursday, July 30, 2009
Friday, July 17, 2009
Video Analysis and Responsive Mirror part II
Responsive Mirror part II
For those who are interested in what happened to the responsive mirror as discussed, here is a look at the second generation of the technology.
It is more integrated in terms of form factor and uses an enhanced video analysis core. As reported by Boing Boing: http://gadgets.boingboing.net/2009/06/30/how-parcs-responsive.html
Video Analysis
For the inner geeks in us, here is a high level description of the video analysis technology used for Responsive Mirror
1. Finding matches across images is performed by deriving a metric of similarity (i.e., a distance) between two images.
2. Then, an optimization is performed to find the best matches across two sets of images that minimize the aggregate distance between the two sets.
Some technical details
1. What is the similarity metric between two individual images?
Modified version of Euclidean distance by treating an image as a large vector of pixel values.
2. What is the “aggregate distance” between two sets of images (as opposed to two individual images)?
Contact me (email: "yfjuan (at) parc (dot) com") for more information on the answer
3. How is the optimization performed?
Greedy descent algorithm to minimize aggregate distance.
For those who are interested in what happened to the responsive mirror as discussed, here is a look at the second generation of the technology.
It is more integrated in terms of form factor and uses an enhanced video analysis core. As reported by Boing Boing: http://gadgets.boingboing.net/2009/06/30/how-parcs-responsive.html
Video Analysis
For the inner geeks in us, here is a high level description of the video analysis technology used for Responsive Mirror
1. Finding matches across images is performed by deriving a metric of similarity (i.e., a distance) between two images.
2. Then, an optimization is performed to find the best matches across two sets of images that minimize the aggregate distance between the two sets.
Some technical details
1. What is the similarity metric between two individual images?
Modified version of Euclidean distance by treating an image as a large vector of pixel values.
2. What is the “aggregate distance” between two sets of images (as opposed to two individual images)?
Contact me (email: "yfjuan (at) parc (dot) com") for more information on the answer
3. How is the optimization performed?
Greedy descent algorithm to minimize aggregate distance.
Monday, July 13, 2009
Video analysis and shopping
Responsive Mirror
One downside of working at PARC is that a lot of technologies sound so out there or address very specific (fundamental) problems that they don't make for good cocktail tales.
Responsive mirror, on the other hand, uses cool technology in a context that we can all relate to. It lets you do side by side real-time comparisons of clothing with prior pieces as you put on a new one instead of conjuring up the images of what the prior piece look like.
Of course, this is not one of 'em before/after pictures in diet pill infomercials. Responsive mirror follows your actual movement and plays back the same movement with the prior article. In other words, you can do not just "apple-to-apple" comparison but "same-side-of-apple-to-same-side-of-apple" comparisons. Now, that is cool!
Here is a link on the tool: http://blogs.zdnet.com/BTL/?p=21045
Video Analysis
Of course, under the hood, PARC has built up a platform on a robust video analysis and given the proliferation of both cameras and screens beyond laptop computers (think smart phones), there are a lot of applications that are still waiting to be imagined.
===
P@P
One downside of working at PARC is that a lot of technologies sound so out there or address very specific (fundamental) problems that they don't make for good cocktail tales.
Responsive mirror, on the other hand, uses cool technology in a context that we can all relate to. It lets you do side by side real-time comparisons of clothing with prior pieces as you put on a new one instead of conjuring up the images of what the prior piece look like.
Of course, this is not one of 'em before/after pictures in diet pill infomercials. Responsive mirror follows your actual movement and plays back the same movement with the prior article. In other words, you can do not just "apple-to-apple" comparison but "same-side-of-apple-to-same-side-of-apple" comparisons. Now, that is cool!
Here is a link on the tool: http://blogs.zdnet.com/BTL/?p=21045
Video Analysis
Of course, under the hood, PARC has built up a platform on a robust video analysis and given the proliferation of both cameras and screens beyond laptop computers (think smart phones), there are a lot of applications that are still waiting to be imagined.
===
P@P
Subscribe to:
Posts (Atom)
Posts
