[Rock-dev] [Discussion] usage of infinity for "no-data" tagging

Jakob Schwendner jakob.schwendner at dfki.de
Mon Aug 1 09:43:53 CEST 2011


On 07/29/2011 04:31 PM, Alexander Duda wrote:
> On Fri, 2011-07-29 at 09:58 +0200, Jakob Schwendner wrote:
>> On 07/28/2011 02:10 PM, Sylvain Joyeux wrote:
>>> I've created a proposal for a guideline page for that
>>>
>>>      http://rock.opendfki.de/wiki/WikiStart/Standards/RG3
>>>
>>> Please discuss. I'll add the relevant bits to base when we agree on the
>>> proposal.
>>>
>> ok for me regarding unset values and the usage of NaN.
>> I don't agree with the usage of infinity though. Infinity has certain
>> properties that do not hold true for an "unknown value".
>> I've done some tests on inf, because I wasn't 100% sure if my
>> assumptions hold, but they do:
>> a : 1
>> inf : inf
>> minf : -inf
>>
>> a<  inf : 1
>> a>  inf : 0
>> a<  minf : 0
>> a>  minf : 1
>> a / inf : 0
>> a / minf : -0
>> a * inf : inf
>> a * minf : -inf
>> inf * -5.0 : -inf
>> inf / -5.0 : -inf
>>
>> I agree with the usage of inf in covariance matrices (since it's not an
>> unknown value there), but not the example with the vector. If you really
>> don't want to represent unknown values in a separate variable I would
>> even prefer nan in this case.
>> The thing is you can do normal calculations with inf and even get a real
>> number (0) out of it in case you divide another number by inf. Not what
>> I would expect from an unknown number.
>>
>> Jakob
>>
> For me inf is a sentinel value and cannot be measured by any physical
> sensor. Therefore in this case inf is just a concept to flag something
> like a unknown value.
>
> Alex
>
Ok, this is getting philosophical now. Inf and minus inf have very known 
properties, like e.g. they are always bigger, no matter what you compare 
it to. And you can calculate with it as shown in the examples above. And 
that you cannot measure it is also not true. Let's say you have a 
distance image that you calculated from disparity values. If your 
disparity value is 0, which is not unreasonable given quantization 
involved due to image resolution a.s.o., you get a distance value of inf.
Using inf to flag that a value is unknown is wrong in my opinion. 
Unknown means it could be -4, 1, 10 or 20 or whatever. Yes, they 'value' 
of inf is also unknown, but it has very known properties in relation to 
-4, 1, 10 and 20, as it is always greater. Maybe one could say inf is 
unknown, but unknown is not inf :)

Jakob

-- 
Jakob Schwendner, M.Sc.
Researcher

DFKI Bremen
Robotics Innovation Center
Robert-Hooke-Straße 5
28359 Bremen, Germany

Phone: +49 (0)421 17845-4120
Fax:   +49 (0)421 17845-4150
E-Mail: jakob.schwendner at dfki.de

Weitere Informationen: http://www.dfki.de/robotik
-----------------------------------------------------------------------
Deutsches Forschungszentrum fuer Kuenstliche Intelligenz GmbH
Firmensitz: Trippstadter Straße 122, D-67663 Kaiserslautern
Geschaeftsfuehrung: Prof. Dr. Dr. h.c. mult. Wolfgang Wahlster
(Vorsitzender) Dr. Walter Olthoff
Vorsitzender des Aufsichtsrats: Prof. Dr. h.c. Hans A. Aukes
Amtsgericht Kaiserslautern, HRB 2313
Sitz der Gesellschaft: Kaiserslautern (HRB 2313)
USt-Id.Nr.:    DE 148646973
Steuernummer:  19/673/0060/3
-----------------------------------------------------------------------




More information about the Rock-dev mailing list