The first light field camera for the consumer market
Todor Georgiev and Andrew Lumsdaine
Lytro
Ren’s Dissertation 2006 “Refocusing”
Levoy-Hanrahan “Light Field” (1996), Gortler et al. “Lumigraph” (1996), Adelson’s “Plenoptic Function” (1991), etc.
Lippmann’s work on capturing radiation with array of lenslets “Integral Photography” (1908). Nobel Prize (Color Photography)
Physical quantity: Radiance = Energy density in 4D ray space.
Origins
11-2
Lytro Plenoptic Camera
Existing commercial platform for characterizing problem space and for new algorithm development and exploration
Company founded in 2006 (as “Refocus Imaging”) to commercialize Ng’s PhD thesis work at Stanford – handheld plenoptic camera
First camera available for sale October 2011 11Mpx CCD captures 11 “megarays” Postprocessing accomplished on host
Originally Mac only Now Mac + PC
Creates focal stack of images Illusion of real time refocusing New effects released Dec 4 2012
Perspective Instagram-like effects
Implements microlens array approach to plenoptic imaging Lytro sensor: 1.4 micron pixels with 14 micron microlenses in
hex array
Basic Lytro design: Microlens array
Microlens arraysensor
8X optical zoomOptical focus
Analysis of traditional camera imaging
The outside 3D world is mapped into the inside 3D world. Projective transform mapping points to points, lines to lines, planes to planes. Infinity treated projectivly. Points, lines and plane at infinity handled seemlesly.
Analysis of traditional camera imaging
The image plane (sensor) captures sharp all points that happen to be mapped to its location. Everything else has certain amount of blur. This is based on the mapping of rays to rays, points being defined as the vertexes of pencils of rays.
Conventional camera image
In a conventional cameraonly the area around the image plane is in focus (DOF). The rest is blurry.
Out of focus
In focus
Out of focus
Analysis of plenoptic camera imaging
If pixels are replaced by microlenses positioned at distance f from the sensor, ray intensities would be directly recorded. Thus the plenoptic camera captures the 4D image of ray intensities (the radiance), and not a 2D image. Full record of the radiance of a scene.
Analysis of plenoptic camera imaging
This approach (pixels replaced by microlenses) for recording ray direction produces 1 pixel per microlens, which would be 0.1 megapixels for Lytro. Our MTF measurements show 0.3 megapixels and in certain cases even higher resolution. How is that possible? (see next)
The plenoptic camera as a relay system. Shaded area represents area of good focusing of the microlenses (at Nyquist). In that area we can have full sensor resolution rendering from each microimage. Mixing such microimages produces the high final resolution that we observe. We call this “full resolution rendering.” The unshaded area can render only 1 pixel per microimage, and is inside the hyperfocal distance f ² / p from the microlenses, where p is pixel size. (This is approximately 0.5 mm in Lytro) Georgiev, T., Lumsdaine, A., Depth of Field in Plenoptic Cameras, Eurographics 2009.
Analysis of plenoptic camera imaging
In a plenoptic camera DOF is extended, but the central part can never be recovered in focus from individual microimages
In focus
Out of focus
In focusThis result is based on our camera similar to Lytro:Georgiev, T., Lumsdaine, A., Depth of Field in Plenoptic Cameras, Eurographics 2009.
Analysis of plenoptic camera imaging
Plenoptic 2.0 camera
The plenoptic 2.0 camera as a relay system. Shaded area represents good focusing of the microlenses, satisfying the lens equation. In that area we can have full resolution rendering and super resolution that can be 4X better. The unshaded area should be excluded. This approach is good for image capture close to the microlenses, but it has lower DOF. Used by Raytrix.Lumsdaine, A., Georgiev, T., The Focused Plenoptic Camera., ICCP 2009
Lytro: The Captured Image
Lytro: The Rendered Image
Lytro: The Rendered Image
More technical detail
Lytro: More technical detail
Lightfield Data for Algorithm Development
Lytro application stores three main sets of data (organized in sqlite db) Camera calibration data / modulation images Raw lightfield files Processed lightfield files (focal stacks) are computed locally and
stored Raw lightfield files
Not demosaiced Some meta information about the shot JSON header plus raw 12-bit data
The RAW Microimages show vignetting, noise, random shift of microlenses, etc. To correct, a calibration step is required as imperfections are camera specific.
Modulation images are included with each Lytro camera (12bit images with time stamp). Calibration images summary: • 60 modulation images are captured for each camera at manufacture
time (30 min based on file time stamp). Different lens settings, like focus, zoom, exposure.
• Two dark images at different exposure. Our modulation images usage for Lytro rendering:
• Divide the captured image by the corresponding modulation image (anti-vignetting) at similar parameters. Clean up pattern noise, dark noise.
• Compute the true microimage centers. Use the new centers for rendering. This is the most important calibration in our experience.
• Possibly Lytro is using lens model to compute centers.
Factory Calibration
Lytro: Modulation images
Lytro: Modulation images
"clock": { "zuluTime": "2012-03-27T05:24:30.000Z" }, "pixelPitch": 0.000001399999976158141876680929 }, "lens": { "infinityLambda": 7.0, "focalLength": 0.05131999969482421875, "zoomStep": 100, "focusStep": 832, "fNumber": 2.21000003814697265625, "temperature": 38.569305419921875, "temperatureAdc": 2504, "zoomStepperOffset": 2, "focusStepperOffset": -36,
Lytro metadata examples
"mla": { "tiling": "hexUniformRowMajor", "lensPitch": 0.00001389861488342285067432158, "rotation": -0.002579216146841645240783691406, "defectArray": [], "scaleFactor": { "x": 1.0, "y": 1.00024712085723876953125 },
So for example, microlens pitch is 13.9um, and the microlens array is estimated to have rotation angle -0.00258 relative to the sensor. Zoom step and focus step change for each picture. Our calibration is done by trying to match the image parameters with calibration images having closest metadata.
Demo of rendering Lytro
Demo of rendering Lytro
Lytro MTF: Target 15 and 20cm from the camera
15cm
20cm
Compare with halftone printing by dithering
This effect is characteristic for 1.0 camera at the depth corresponding to the microlenses. It’s the price we pay for extending depth of field / good refocusability.
Real world Lytro example
Real world Lytro example
Zoomed in refocusing
Zoomed in refocusing – note the dot artifacts
Microimages of constant color
Conclusion: Lytro and resolution of plenoptic cameras
Conclusion:
Lytro is the first light field camera for the consumer market.
It’s likely that Lytro renders images based on a version of the full resolution method, generating much more than 1 pixel per microlens. The Lytro camera and application appear to reproduce sensor resolution captured by each micro image. However due to mixing of multiple views, final image resolution (under 1 megapixel) is far below sensor resolution (11 megarays).
Typical numbers for full resolution rendering from plenoptic camera data are 10X -- 20X less than the sensor resolution. That’s for Lytro and for any other rendering.
This situation can be greatly improved with superresolution. Results with resolution only 5X lower than that of the sensor have been demonstrated.
Lytro too have been able to generate much higher resolution than their current rendering, in certain cases.