Light field toolbox v0.5

[Donald Dansereau]

Hi all,

After a long pause the light field toolbox for matlab is again under active development. The latest release v0.5 is on github and includes:

• Linear refocus super-resolution using LFFiltShiftSum.
• New display functions LFDispLawnmower, LFDispTiles, LFDispTilesSubfigs, LFDispProj, LFDispProjSubfigs
• LFReadESLF, LFWriteESLF
• Improved decode performance and speed
• Improved calibration accuracy
• LFDisp* functions are better behaved, now display in the active figure window

https://github.com/doda42/LFToolbox

links to datasets and community resources:

http://dgd.vision/Tools/LFToolbox/

New features for v0.6 are already underway, and longer-term more significant improvements to decode and calibration are in the works, including stable calibration for Illum cameras.

Feedback and contributions are welcome, get in touch via this forum or via the github page.

Best regards,

Donald Dansereau

[Yudong Liu]

Hi Donald, 

Thank you for continuing to work on Light Field.

I am doing research using the Light field camera for combustion, as a PhD student. The previous LFToolbox 0.4 really help me out.

Want to say thank you. And also looking forward to the ver. 0.6!

Best,

Yudong

在 2020年5月14日星期四 UTC+8上午9:53:32，Donald Dansereau写道：

在 2020年5月14日星期四 UTC+8上午9:53:32，Donald Dansereau写道：

[Gene LeSage]

I was able to decode the sample images in Matlab light field toolbox (LFT) with identical results as shown in the manual but LFT decoding of images from my camera look noisy compared to same images processed by the Lytro desktop.  I extracted white images from my camera as described in Appendix and ran through all steps of decoding described in manual without any reported errors.  Example of image processed by LFT and Lytro desktop are below.   My camera is F01 with 1.2.2 firmware.  Without reported errors, I am not sure where to begin to find the problem source.    

decoded by matlab FTL

Lytro desktop, far field

Lytro desktop, near field

___________________________________________________________________________

[Donald Dansereau]

FAQ: Why is it so noisy?

There will always be important differences between the Lytro software output and the toolbox output. The toolbox tries to generate a 4D light field that is as close as possible to the raw image measured by the camera, while still being a standard two-plane parameterized 4D light field. The Lytro software has a very different goal. They do not produce a 4D light field, they produce 2D renders. These are optimized to look nice, and evidently look much nicer than any 2D slice taken from the toolbox output. They use sophisticated decoding and denoising techniques to do this.

The philosophy of the toolbox is to provide a 4D light field close to the raw image captured by the camera, to allow researchers to explore the characteristics of this kind of signal. It should, in theory, be possible to go from the 4D light field output by the toolbox to nice 2D renderings like those produced by the Lytro software.

Note that the issue is not only the noise spread over the 4D signal, it's the colour fidelity. The raw sensor readings are pretty far from yielding colours in any standard colour space, as is generally the case for low-cost CMOS image sensors. Fixing this is however complicated by vignetting at the lenslet image edges. There is substantial work ready to be released into the LFToolbox around dealing with this colour correction, I am seeking collaborators or spare minutes to make this happen. It is mostly a documentation and code cleanup effort at this stage, and the code (care of Mikael Le Pendu and Pierre David, who both pioneered the methods and put a substantial effort into putting the code into the toolbox) is awaiting this effort in the develop0.6 branch.

Hope that clarifies things,

Don