Hi Yulia,

I, too, am looking at HDR measurements of artificial light sources in
the hope of measuring glare. The results I've been getting are rather
inconsistent, so I've been doing some googling and reading.

> I was able to figure out luminance values for a single LED, which can be
> compared to the ones from HDR images. But I have a couple of
> questions/concerns on HDRI technique and Photosphere.

> At first, I’ve used “regular” scene to retrieve response curve of the camera
> (large smooth gradients with very dark and bright areas, and had reflectance
> standards for the absolute calibration).
>
> Camera: EOS T1i Rebel with 28-105mm lens, at 28mm
> Calibrated at the grey reflectance sample 186.45 cd/m2
> CF=0.957

> Then I use this RC to analyze HDRI of a captured LED. The value is 230,000
> cd/m2 for a single LED, which is low (it’s has to be around 7*106 cd/m2).
> So, it underestimates the luminance.

McCann and Rizzi have done some quite comprehensive research into the
dynamic range of cameras and how it is limited by veiling glare. They
have published an HDR book
"The Art and Science of HDR Imaging" (Wiley, 2011), and many of their
papers are available on
http://web.mac.com/mccanns/HDR/Glare_Limits_HDRI.html

To wet your appetite, I recommend
http://web.mac.com/mccanns/HDR/Glare_Limits_HDRI_files/07EI%206492-41_1.pdf
Their conclusion is that accurate HDR luminance measurements are not
actually possible because veiling glare that is generated within the
lens limits the dynamic range of the optical system. Apparently, there
is actually an ISO standard (9358:1994) that comes to the same
conclusion: The higher the dynamic range of the scene, the more
inaccurate the HDR measurement.

You will be aware of the 'flare removal' option in hdrgen (-f switch).
I'm not entirely sure where 'flare' sits between 'point spread
function' and 'veiling glare', but I believe it to be closer to the
former. The PSF is a funciton of any optical system that results in
the image to become 'smudged' out. Back a few years ago when the
megapixel race was in full swing, many observers correctly stated that
the lenses on cheap digital cameras can't actually provide a
resolution that a would justify, say, 12MP on a digital snap shot
camera. This is the PSF they were talking about--it's how a pixel
affects the neighbouring pixels.

While the PSF can be estimated (or even calculated, given enough
information about the lenses and their optical properties? Not
sure...), veiling glare, on the other hand, cannot because it depends
on the scene. Every 'pixel' of the scene affects every pixel of the
image. It's even worse than that: Even scene objects outside of the
field of view of the optical system have an impact on the sensor
image.

Hoefflinger (Ed.) "High-Dynamic-Range (HDR) Vision" (Springer, 2007)
has an entire section dedicated to HDR lenses. While true HDR low-res
(video-) cameras are actually becoming commercially available (they
have a logarithmic response, with a dynamic range far exceeding that
of the human vision), the problem is that they require special HDR
lenses that have to be carefully designed to minimise veiling glare.
Digial camera lenses (even pro-level DSLR ones) are not optimised for
this.

So there is nothing wrong with the camera calibration that you carried
out with a LDR scene. This is how it should be done. The problem
you're facing is not specific to Photosphere or the Mitsunaga RSP
recovery algorithm. The RSP is not compressed at the upper end--it's
just Physics that you're up against.

> It seems like calibration point is critical here. I’ve decided to try to
> capture a different scene for deriving RC with a wider range. It would make
> sense that camera has to see higher luminance values in order to accurately
> measure them later. The dynamic range has to cover measured values.
>
> 1.   1. How does Photosphere deals/approximates/calculates the upper end of
> the curve? I assume it gives more weight to mid tone values? But what
> happens with high luminance values?

> 2.     I assumed when CF is applied, it does not equally change all values,
> but does it proportionally to RC (since it is not linear).  Why does it do
> it equally for the whole range?
>
> Lsun=80*106 cd/m2. And of course CF is very big 391.

> 3.     Does photosphere compress the response curve, so at the upper end all
> values above certain threshold will have the same number?
>
> 4.     Any additional suggestions on properly obtaining and calibrating HDRI
> for this purpose?

I'm afraid you have to lower your expectations with regards to the
achievable accuracy when it comes to HDR scenes that include bright
light sources.

Light modulation ('flicker') is another problem with HDR measurements
of electric light sources. Unless you are certain that your light
source is driven by a HF driver or ballast, I recommend you actually
measure the modulation of the light source. If the LEDs are mains
driven, they will flicker with 100 or 120 Hz, depending on your mains
frequency. If the modulation factor is high, e.g. if the LEDs
effectively switch on and off with this frequency, HDR measurements at
short exposure times will be unpredictable. You can test this by
taking a number of photographs of the same scene (with light source in
it) at short exposure times. There is no need to go HDR. If all images
have the same overall 'brightness', you're all right. If the images
are noticeably different, you've got yet another problem.

Cheers

Axel

_______________________________________________
HDRI mailing list
[email protected]
http://www.radiance-online.org/mailman/listinfo/hdri

Reply via email to