Slashdot Mirror
Camera Lets You Shift Focus After Shooting
Zothecula writes "For those of us who grew up with film cameras, even the most basic digital cameras can still seem a little bit magical. The ability to instantly see how your shots turned out, then delete the ones you don't want and manipulate the ones you like, is something we would have killed for. Well, light field cameras could be to today's digital cameras, what digital was to film. Among other things, they allow users to selectively shift focus between various objects in a picture, after it's been taken. While the technology has so far been inaccessible to most of us, that is set to change, with the upcoming release of Lytro's consumer light field camera."
Enhance.
how is babby formed?
From the sound of it, it basically sounds like it captures a picture with a Z-buffer -- that is, they capture spatial information and angular information, and the angular information is then matched up to find corresponding objects to assess depth for refocusing.
One nifty thing about pictures and videos with built-in Z-buffers would be that it'd be really easy to render into them. Heck, you could have a camera with a built-in GPU that could do it in realtime as you're recording. :)
One step beyond the Z-buffer would be to then do a reverse perspective transformation and extract polygonal information from the scene. This would be of particular use in video recording, where people moving allows the camera to see what's behind them, hidden sides of their bodies, etc. Then you could not only refocus your image, but outright move the camera around in the scene. Of course, if we get to that point, then we'll start seeing increasing demand for cameras that always capture 360-degree panoramas. Combine this with built-in GPS and timestamping and auto-networking of images (within whatever privacy constraints are specified by the camera's owners), and the meshes captured from different angles by people who don't even know each other could be merged into a more complete scene. In busy areas, you could have a full 3d recreation of said area at any point in time. :) "Let's do a flyover along this path in Times Square on this date at this time..."
"99 dead duelists of Dios on the wall. 99 dead duelists of Dios! Take one's ring, pass it around..."
No. This is known as plenoptic imaging, and the basic idea behind it is to use an array of microlenses positioned at the image plane, which causes the underlying group of pixels for a given microlens to "see" a different portion of the scene, much in the way that an insect's compound eyes work. Using some mathematics, you can then reconstruct the full scene over a range of focusing distances.
The problem with this approach, which many astute photographers pointed out when we read the original research paper on the topic (authored by the same guy running this company), is that it requires an imaging sensor with extremely high pixel density, yet the resulting images have relatively low resolution. This is because you are essentially splitting up the light coming through the main lens into many, many smaller images which tile the sensor. So you might need, say, a 500-megapixel sensor to capture a 5-megapixel plenoptic image.
Although Canon last year announced the development of a prototype 120-megapixel APS-H image sensor (with a pixel density rivaling that of recent digital compact point-and-shoot cameras, just on a wafer about 20x the area), it is clear that we are nowhere near the densities required to achieve satisfactory results with light field imaging. Furthermore, you cannot increase pixel density indefinitely, because the pixels obviously cannot be made smaller than the wavelength of the light it is intended to capture. And even if you could approach this theoretical limit, you would have significant obstacles to overcome, such as maintaining acceptable noise and dynamic range performance, as well as the processing power needed to record and store that much data. On top of that, there are optical constraints--the system would be limited to relatively slow f-numbers. It would not work for, say, f/2 or faster, due to the structure of the microlenses.
In summary, this is more or less some clever marketing and selective advertisement to increase the hype over the idea. In practice, any such camera would have extremely low resolution by today's standards. The prototype that the paper's author made had a resolution that was a fraction of that of a typical webcam; a production model is extremely unlikely to achieve better than 1-2 megapixel resolution.
The website about the camera doesn't have enough details, either, but this paper does give a reasonable idea of what's going on.
... demonstated to be a working principle.
The paper includes graphics and formulas... a fuck load more detail than the story link given to us...
For large sets, this will be our guide even unto death, for the LORD will work for each type of data it is applied to...
It's called a Plenoptic Camera. You put a bunch of microlenses on top of a regular sensor. Each lens is the equivalent of a single 2D image pixel, but the many sensor pixels under it capture several variations of that pixel in the light field. Then you can apply different mapping algorithms to go from that sub-array to the final pixel, refocusing the image, changing the perspective slightly, etc. So color-wise it's just a regular camera. What you get is an extra two spatial dimensions (the image contains 4 dimensions of information instead of 2).
Of course, the drawback is that you lose a lot of spatial resolution since you're dividing down the sensor resolution by a constant. I doubt they can do anything interesting with less than 6x5 pixels per lens, so a 25 megapixel camera suddenly takes 1 megapixel images at best. The Wiki article does mention a new trick that overcomes this to some extent though, so I'm not sure what the final product will be capable of.
There has been a fair amount of computer science research over the last decade over what you could do if you took a picture with a plane of cameras instead of just one or two. The resulting dataset is called a "light field". You can re-composite the pixels to change depth of focus, look around or through occluding obstacles, dynamically change point of view, etc. As digital webcams became dirt cheap people started building these hyper-cameras and experimenting with them. people learned you could relatively interesting things with small arrays of 4 or 5 squared cameras. Later on they discovered you do this with one camera, with a multi-part lense, then reconfigure the output pixels in the computer in real time. I've seen all these systems demo'ed at SIGGRAPH over the years. Now someone appears to be commercializing one.
I think the infamous bullet-dodging scene in the first Matrix movie was a type of hyper-stereo camera, a row of them albeit. The output lightfield was reconfigured expand point-of-view into time.
Read this paper (or at least skim it) - these are called plenoptic cameras.
It doesn't do any particular voodoo. I suppose you could distill it down to the point where the camera is (in function) a compound eye.
For large sets, this will be our guide even unto death, for the LORD will work for each type of data it is applied to...
The sacrifice of resolution isn't really that big a concern. Consumer cameras have far more resolution than they need these days, as the almighty megapixel has been used as a marketing ploy even though increasing pixel density on the CCDs has led to lower image quality overall. My 10 year old 2Mpx Canon still takes better pictures than any of my wife's last 3 compact cameras (4, 5 and 8Mpx Nikon and Canons), especially in low light. I would go so far as to say it doesn't make sense to have go beyond much more than 4Mpx with lenses the size of compact cameras, as details will be lost due to lens quality long before the pixel count causes loss of detail.