Floating Point data from the tomograms

[sgvdan]

First of all, thank you for this incredibly useful contribution!

Going in to how the e2e tomogram data is parsed, the bitbucket explanation, talks about the data being an unsigned float with 6 bits for exponent and 10 bits for the mantissa.
However, the code here parses it as a 16-bit unsigned-int and then performs a somewhat weird mathematical manipulation to it -

From my small side-note calculation, this manipulation doesn't coincide with float structure -

Would be grateful for some clarification here :)

Thanks!
Dan

[Dbrown411]

Hi Dan,

You'll want to take a look at the look up table LUT. The bytes are first parsed as unsigned integers (for speed), and then converted to the nonstandard float type via the LUT.

The calculations in the code snippet you posted are for applying gamma to the intensities. The 1/2.4 gamma is pretty arbitrary as far as we can tell.

Dillon

[sgvdan]

I did elegantly ignore that part. Will have a look. Thanks!

[marksgraham]

Yes thats right - I had initially implemented the conversion this way

OCT-Converter/oct_converter/readers/e2e.py

Line 264 in 3063aa8

def read_custom_float(self, bytes):

and Dillon's implementation gives the same results + a huge speed up

[sgvdan]

Indeed a very neat thing to do. Well done!

There is one thing I'm still wondering about: In the read_custom_float implementation, you've subtracted 63 from the exponent. However, according to this page you are supposed to subtract only . Anything I'm missing here?

[marksgraham]

I lifted the -63 directly from the uocte implementation but I don't know where they got it from.

[Dbrown411]

I don't have time to look into this more for a while, and I have a pretty limited background in floating point. But it's my understanding that the typical bias (2^(6-1) - 1) allows for a range of ~(2^-31 - 2^31). So subtracting 63 instead should essentially normalize the pixel intensities to between 0 and 1 (2^-63 - 2^0)?

Not sure what advantages/effects this has on the distribution of pixel intensities (if any). But maybe this has something to do with the seemingly arbitrary 1/2.4 gamma transformation?

[sgvdan]

It basically means that the values we get are all multiplied by 2^(-32).
It is merely a scale factor - that's why I think we haven't noticed it yet. Therefore not a big deal, but indeed a bit weird.