ATDO1 model for color appearances, color differences and chromatic

adaptation
S. Lee Guth, Dept. of Psychology and School of Optometry, Indiana University

ABSTRACT

This paper reviews and critiques earlier version of the ATD model, and it describes a revised model, which differs mainly in
its gain control function. The model is applied to recently published lightness, hue and colorfulness data.

Keywords: Color models, lightness, hue, colorfulness, chroma, saturation color differences, chromatic adaptation.

1. INTRODUCTION
In 195 1 , in a brilliant chapter on color vision, Judd' wrote, "By far, the most completely worked out of these stage theories is
that of Mueller [from 1930]. It takes at least qualitative account of nearly every known [color] visual phenomenon .. .and
he has . . .made a start toward the solution of important problems that will eventually have to be faced by other theorists" (p.
836). Indeed, the ATD models for color perception and visual adaptation, which account for a very wide range of data of
human color vision, suggest that Judd was correct in his judgment. That is, with some changes, the various versions of the
ATD model that have been published during the past ten years2 are quite similar to the Mueller model. Differences include,
(i) changes of the receptor inputs to post-receptor mechanisms, (ii) addition of nonlinear gain control at the receptor level,
(iii) changes of some post-receptor connections and weighting factors, and (iv) addition of other system nonlinearities.

The intent of this paper is to present an up-to-date version of the ATD model, to correct errors that were associated with prior
publications, and to make explicit the formulations that should be used when applying ATD to current problems concerning
color appearances. The paper will also emphasize the fact that the ATD models remain extremely powerful for solving the
classical problems of color discriminations and chromatic adaptation.

2. DESCRIPTION OF THE MODEL, CHANGES, PRIOR ERRORS AND COMMENTS

Figure 1 schematizes the model. At the first level, the model uses generally accepted LMS receptors. Nonlinear (NL) signals
from receptors are affected by nonlinear gain control mechanisms, which allow maximum neural information to flow at low
light levels but help to prevent saturation at high levels. Gain control is also affected by neural activity from other retinal
areas and from prior retinal stimulation. These gain control effects allow the prediction of simultaneous and successive
chromatic adaptation data. The gain-controlled outputs of the receptors then feed a lower-level post receptor stage that
consists of two "opponent" mechanisms (T1 and D1) and one additive nonopponent mechanism (A1). (In ATDO1, the value for
C2 15 zero.) Neural activities from these three mechanisms are compressed nonlinearly and then define a three-dimensional
Euclidian vector space. The length ofA1T1D1 vectors in that space relate to the apparent brightnesses of light, and the
distances between the vectors relate to discriminations between and among chromatic or achromatic lights or
samples. The length of the A1 vector alone is very nearly proportional to CIE luminance. (Due to nonlinearities, A1
looses strict proportionality as luminance increases, but note that changes in A1 mimic level-dependent flicker-photometric
spectral sensitivity functions that have been reported.) Outputs from the first post-receptor stage feed a second stage that is
(after compression) a classical opponent colors stage. The A2 mechanism signals an achromatic percept, the T2
mechanism signals redness (+T2) or greenness (-T2) and the D2 mechanism signals yellowness (+D2) or blueness (-D2).

The ATDO1 model incorporates some changes that were made in earlier versions, but it also differs in several ways. (i) A
slight and probably inconsequential change was made, and maintained here, in the equation for the S receptor. (ii) For ease of
calculation, Judd's3 "corrected" XYZ functions are not used (but, they should be used if one wants luminances to be based on
the "corrected' spectral sensitivity function, and/or if near-spectral lights in short wavelengths are of concern). (iii) For
suprathreshold predictions, receptor "dark light" or "noise" is neglected. (If noise is required, the S receptor's high noise in
ATD952 should not be used, because it causes a severe shift in the model's white point at low levels.) (iv) To conform to
common practice, the signs for yellow and blue are reversed relative to earlier models. (v) The various weighting coefficients
are changed, and (vi) a new gain control equation is used. Items (v) and (vi) are discussed in the following paragraphs.

9th Congress of the International Colour Association, Proceedings

of SPIE Vol. 4421 (2002) © 2002 SPIE · 0277-786X/02/$15.00 303

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/19/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 Stage 2 Correlates An explanation for the weighting coefficients in ATDO1
Stage 1 Correlates
raises profound issues regarding white-point normalization
in color appearance experiments. The ATD models have
been criticized, because they do not allow such
normalization, but that is strength of the model. All other
models, because of inherent limitations, must normalize
white points, and experiments designed to test those models
implicitly recognize the limitation by defining the neutral
sample. However, ATD models make absolute, not relative
predictions. An ATD model can be thought of as an
"Average Observer," who has his/her own white point, but
that theoretical observer cannot be persuaded that a non-
neutral sample is really achromatic. Within the framework
of ATD, all experiments should be carried out without
normalization of the neutral point. Other experimental
data are now obsolete (at least for model competitions)
because they were conceived within an incomplete
theoretical framework. Observers should be required to
make their judgments just as they perceive the samples --
the model's gain control mechanism automatically adjusts
for changes in adaptation produced by varying the
illuminants, because it is, by its very nature, a model for
chromatic adaptation. To apply ATD to appearance data
its neutral point is moved by changing the weightings for
the receptor inputs to the first opponent stage. However,
this is not satisfactory, because the model will then "behave'
as if it has that white point, whereas real observers will still
judge most samples based upon their (the observers) true
Figure 1. Diagram of ATDO1
achromatic points. Therefore, after adjusting ATD's neutral point to conform to the normalization of a particular experiment,
further fine-tuning of the receptor weights is required. Note that ATDO1 is not to be considered as a normative model for
an Average Observer-- rather it is normalized and adjusted to conform of the normalization of the data being
predicted. Previous models are closer to a normative Average Observer.
2.1 The new gain control equation
The gain control equation for prior ATD models was inadequate for two reasons. As is true for ATDO1, predictions of
adaptation effects assume that receptor gain control is affected by adapting fields or lights of various sizes and durations, and
the model includes an adjustable weighting factor for variations of those influences. For example, if a large red field exactly
surrounds a white target, then the gain control effect of that red light might be multiplied by, perhaps, 50, and the resulting
predictions of a strong dimming and greening of the target were excellent. A logical problem was that the same rule would
apply if test and adaptation fields were identical, and this incorrectly implied large changes in target brightness, if the target's
size were merely increased. A second problem concerned the fact that a given adapting field will not have the same
proportional effect on a slightly dimmer field than on a much dimmer field. A model should not suggest that the attenuating
effect of an adapting field can be expressed as a multiplicative reduction factor. Note that this has nothing to do with the
nonlinearities in the system -- the problem exists no matter what nonlinearities lead to the attenuating factor. This problem
existed in prior ATD models, and it meant that ATD could not possibly have predicted much of the Werner and Walraven4
data. Indeed, an extremely unlikely combination of two programming errors made impossible predictions look very good
(and this author very bad). Both problems suggest that gain control must depend not only on a receptor's own activity
(self adaptation) but also on the difference between receptor activities for a target light and an adaptation light. The
following gain control equation (for L as an example) incorporates this idea. (L is for the target and La for the adapting light.)
.
Attenuation FactOr = o/[ a+ L + k (La - L)], onlr if La > L, otherwise (La - L) = 0
The value for a is fixed in the model, and k depends upon the size and duration of adapting lights or fields. This equation reduces
to self adaptation when the adapting receptor activity is equal to, or less than, receptor activity from the target, and it means that,
for a given adaptation light, the attenuation factor increases as the target light dims.

304 Proc. SPIE Vol. 4421

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/19/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 JOCIY SNOLLVI1O
J SI pOUIflSSIV'flZLVSflJflUIflSU 'SO1lj1A 1P!M ULJ 3uiójoqU 'Spuqoi3 m po!Jpods ioj i uoti 11iom iij2ij Jo ojduws 05W
pogioods snui oq oqi uottmdupi 'snji y ' ' qoiqtt qiui oq posixj uo ouros poiqpn uotniiquioo jo u uqivoid
'sPP1 OtT.L uiojqoid Si 01 OA!1op PU!J SORjEA JoJ Ij'rv pu '(ci pu1 cl'zv Pn •z:7 3JOfJJ 'Wqi ji 3j St o3u1tI!WnJ w
P
'P1i01 J! A S! U! 'W/pO 1101j3 = { (O 9A
(jr) swuodsaiw gjj7 spuuinp .ioj qoq '7) 'J4J Pu (s pUR 1pBppij "N Pu •4:"S
'ii = '0)X91+ A9O OLO00 vr = xot'o) + .1 A91+ OLOO pu s = +uioo) OLO°
= 'o)X9J + 9ç.Ø OL.OE(ez17:O.O = 1xot.o-) + 1 A91 + tRO.O OLO' PW S= +BALJO*O)QLOE(zLro

(zt) JR1II3I '7 J4J pui 'g II3!M ai aodai UodSU J3J1 U!E IOJP'03 iti gjijj ao;da'sii an Plld3Inw
'ç i
£q MJdSaI UO4IflUfl1 'SIO3IJ f(77)j+7-oJ/D 10
pui £pto (i<"7) o (W(VFV) Jo (s<0s) '
[4ji+NJ/ '[()ij-//D uAt =° o pu,
j
'sMflqJ (,J) (jij"jr) (s-"s)ao
JO
= •0
= 'E(iri)cc+1+ooz)/ooz11 = '1(1'\T-'N)c+J'1+oob/oo1N s = (s-'s)cc+s+ooz)/ooz}s '1-'}P S ZT {AoqI
(r€) lBU! 4Dsqns) (! passaidwonu, $suod3,1 .ioj q1
psaidwoaunSUOdaI .ioj qj a3s •swstwq33w
$ I 'swS!uIuPw PUl q4
-
dsqns) (I
='V ÷'TLcE t'9Z =L 069 'I 06•9 yT =ItcI £:8O +'1 001 s
!LV6OO=V co÷!iooi IIU_=!zcL
() = vuij (pwsaidwo)'j'iofuodsai pua2gvoj
9tcE

"V / ooz) + (I"vI 'i = "i / ooz)+ (Ujj /

'a = "ci oo) + (I"d
'yI+ooz)/v=Zy '(+z)/L=L (u1+ooz)/a=a
1' 1LV'IIDWDAUYWWI1SSU1SVIWIO1 IO1ODS33JVIVIdJVLMV UO1O SDM31AdW

uoinu (vg)
si si mjnin uosod m
o j 'dsjj t
twpi o oqi ssouiipq jo oouoipi (01) pojos o
s
(i) ouj sujqki(.i) ijo idums Si juoiodoido 'tpuojioiot pui oqi
j suqfl
(j7) ijo so ojduws (j) Si SI! ssou1qIq
(n) 'Apiuoj SOjdtU1S ssupipojo '(oi) woiq ('j) UB
poopoid uisfl 3tfl OIU SO 3flIUIO1t01 opituoiqo
ipt suewnb tv-I 'uioq 'MOfiOJpfl sis
(!!') oqL OWjoJ100 flflJO(,j) Joj ojduis i
'onjjuooi we •pGapnIa
= +'Y) +ztL(z' 'oco 1n = trJ r'IEI '001 oc = = rs = +zJ) Z/0ç0 1W (z
H = iiiy /Z[)'(j 'soop qt uoiuow oi oqj imJpI?nbin qoqi olil oiw (sjpj
uioj) oozo
upiu2o4 iooo 'yip t
jj 'Ir '1 pij sonp om pouuiipp JOJ OM 'sqq tflt P "!°1 °'U OOUOJOU!P uooMoq
211!PU0dSU0QftçI SOt1jCAUOlfl Olfl AUWJOS UOOMlOq Oqi OM S11U ! PJ1JU! icq 2uwjnojio oqi '°1°-'°1"! 'oous!p
qoiqt si oq djs-flhiii .iooz 'ouaiajpp p'
iyj jflOM oq Ajowixoidde oonoo ioj utppoidsuwpy 'SOSdqp StOItjM ç
oin oouop pOJflb31 JOJ uofltuTUI!iosIp 'Sp4IOdO 'OSJflOOJO UO J(UtLu -siopj V
sinj Uq piinoj o oq 010W
owidoiddioj UPIJXUd E OpiAt A3OflA JOIflOJO UOWUUIJOSp j1IB UOflOOIOp sjnso1 siuouipnr :iiijo JOO3's3uaEW!p "dv
oxe Ajaup O11jOJO JO1OoA1ou! SOUIJSIpU! •ooaIszUzLzV

'y) + z:'IV + ('UV Ij

=
' oco

DI1VWOfflD WOILVJ4IVUV
PIW +VV)= +LV JV •oco

oqi ndu 'snpt '

Acj uoqi Aiot 'samu jp jjy sjoxxu oq uooq oijiwoitp uoudtp'sppou' i"q TOLLV S3jUIWtfl
pim oi i
ioj ouioitp uijdipi
owidoddeooiotp (yjo oidgomuijuict suosnjj
qj'q
io 'ppguq iociv ppid
UOAO 1QW .SllOFtqO JI
tt
TWuJOi
nq imiioduqsuoipqid3fl UMOS U! •spflj
uI uo!w q)
sno
9 'SMOIJ3I11JNOISSI)JSIU IMV A1VhM11S
çtWflf puqo 'ssotpq2q 'mj ssunjiojoo pm uoivnsSiU3Wpflf
'oi '9 q v Jo
z;ci 2UI3ffOJ Sj4UUS tLL szjj: psn JoJ upi
SflJ1?A it
jmy oppaid 'z="x ='A pu 'nr="z
powjnojuo ioj ipo OjdUICS TI!stt SUOflbO wag SUOPOS
swuipnf I? UOPUIIJ JO 3UIU!WIlj 1ON ddn iqp
m sptmjon ioj oq 'suoipid imiij 'ailly pu1rc1t1v

Si ou e 1
O 'fl AOq Jfl!d s2ttoqs stmnç pouiqo ssuq2J
jrnod nq 1!WJ susicbi Ifl p31qpp !
Ij
SSU14 JO LD o3uaIGJJ 1th1 JOd uJos 'uos1oJ SJdUUS JO tUCS '1flLflWflj 10 iflEtUOSSG 31fl UI1S 'oOtnu!wni m

Proc. SPIE Vol. 4421 305

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/19/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 reference were judged less light than the reference; therefore, for prediction purposes, the sample judged to have maximum
lightness was the reference. Figure 2 shows the predictions. Figure 3 shows hue angle predictions, which can be grossly
deceiving when even miniscule deviations are either across the O- 3600 boundary, or around the origin. Fig. 4 shows colorfulness
predictions. It might be surprising that the best colorfulness predictions were made using what is defined as saturation within
ATD models, and what others might call chroma. That is, for predictions, ATDO1 forms the ratios ofsamples' chromatic signals to
their achromatic signals, and, because the achromatic signals are proportional to luminance, this is equivalent to comparing the
samples to equal luminance white references. (For carefully considered reasons, CIE definitions are rejected here.) Although not
shown, ATDO1's saturation correlate also predicted colorfulness judgments from the only set of LUTCHI data (R-HL, phase 1)
that was examined. Figure 5 shows more about the saturation-colorfulness-chroma question. It compares the observers'
colorfulness and saturation judgments. Note that many samples (which were all dark samples) with near-zero colorfulness were
judged to have almost maximum saturation. This did not seem reasonable, especially because many of the high saturation
samples were near-neutral. Therefore 30 samples with luminances below 20 (max = 100) were eliminated from the comparison,
and Fig. 6 shows the result. Because what others call "chroma" made the best predictions of the colorfulness data, and because of
the correspondence between colorfulness and saturation shown in Fig. 6, it appears that judgments of colorfulness, saturation and
chroma all depend upon the ratio of the chromatic to achromatic components of an observer's percept.
In conclusion, consider that ATh is not a model in the usual sense. It is best described as a quantitative theory of color vision that
builds on and validates the theoretical ideas of our greatest contributors, including Helmholtz, Hering, von Kries, and Mueller. It
is very powerful, and it should now be seriously considered by the vision community as a replacement for all models that are
currently used to make predictions (or to establish standards) that concern human color perception.
12U 120
. 400 ——-
________________________
4
•
,1oo 1oo 4'

::

I 4944
jE ___________
50 100 150 0 100 200 300 400
0 200 400 600 800
Lu.inrice Predicted Lightne5 Predicted Hue Angle
4 Fiqure I . Luminance vs Iiqhtness • Figure 2. ATDOI lightness predictions • Figure 3. ATDO1 hue predictions
— u
. 60 .
.
80
4 4 •• +j*

IE
0.0
RO71
Otuldul •
1.0
.
2.0
.

a.o 4.0
I-, E
0
**
o ioo
IE j:
-
0
o o ioo
Predicted Colorfulnee Judged Colodulne Judged Celoifulnes
4 Figure 4. ATDOI colorfulness predictions • Figure 5. Colorfulness vs saturation • Figure 6. As fig 5, w/o dark samples

7. REFERENCES
1. D. Judd, "Correlates of the visual stimulus," In Stevens, Handbook ofExperimental Psychology, Wiley, New York, 1951.
2. S. Guth. See references in three articles reprinted in R. Eschbach and K.Braun, Recent Progress in Color Science, Society
for Imaging Science and Technology, Springfield, Virginia, 1997.
3. D. Judd, "Colorimetry and artificial daylight, " Proc. CIE, part 7, p. 1 1, Paris, 1951.
4. J. Werner and J. Walraven, "Effect of chromatic adaptation . . .background color," Vision Res. 22, pp. 929-943, 1982.
5. G. Juan, Verification of Colour Appearance Models Using Magnitude Estimation Data, Ph.D. thesis, University of Derby,
Derby, UK, 2000.
* Correspondence: email [email protected] for information or current address.

306 Proc. SPIE Vol. 4421

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/19/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx