Vision Research 40 (2000) 745 758

www.elsevier.com/locate/visres

Spatial resolution and receptive field height of motion sensors in

human vision
Mark A. Georgeson *, Nicholas E. Scott-Samuel 1
School of Psychology, Uni6ersity of Birmingham, Birmingham B15 2TT, UK
Received 1 June 1999; received in revised form 5 November 1999

Abstract
We estimated the length of motion-detecting receptive fields in human vision by measuring direction discrimination for three
novel stimuli. The motion sequences contained either (i) alternate frames of two differently oriented sinusoidal gratings; (ii)
alternate frames of vertical grating and plaid stimuli or (iii) a vertical grating divided into horizontal strips of equal height, where
alternate strips moved leftward and rightward. All three stimulus sequences had a similar appearance (of moving strips) and the
task was to identify the direction of the central strip. For sequences (ii) and (iii), performance fell as the strip height decreased.
Threshold height fell with increasing contrast up to about 20%, then levelled off at the critical strip height. Temporal frequency
(1.915 Hz) had no effect on the critical strip height. We argue that the receptive field length corresponds to twice this critical
height. The length estimates were strikingly short, ranging from about 0.4 cycles at 3.0 cpd to 0.1 cycles at 0.1 cpd. These lengths
agree well with the estimates derived at threshold by Anderson and Burr (1991, J. Opt. Soc. Am. A8, 1330 1339), and imply that
the motion-sensing filters have very broad orientation tuning. These and other results are interpreted within the framework of a
Gaussian derivative model for motion filtering. The sensitivity of motion filters to a broad range of orientations suggests a simpler
view of how coherent plaid motion is processed. 2000 Elsevier Science Ltd. All rights reserved.
Keywords: Psychophysics; Motion detection; Receptive fields; Direction selectivity; Human vision; Gratings; Contrast; Direction discrimination;
Gaussian derivative; Spatio-temporal filtering

1. Introduction
In natural vision, the speed and direction of motion
in the visual field vary with spatial position, and these
spatial variations carry rich information about the spatial structure of the world and about the observers own
movements. In motion analysis there must be a tradeoff between the requirements for robust representation
of local velocity and for resolution of changes across
space. Integrating information over a relatively large
region should improve the reliability of velocity estimation, but at the expense of smoothing out local variations that may be important (Braddick, 1993). While
vision may handle this problem at several levels of the
motion pathway, the width and height of early, motion* Corresponding author. Fax: +44-121-4144897.
E-mail addresses: [email protected] (M.A. Georgeson),
[email protected] (N.E. Scott-Samuel)
1
Present address: McGill Vision Research, 687 Pine Avenue West,
Montreal, Quebec, Canada, H3A 1A1.

selective receptive fields must impose a limit on the

spatial resolution of motion signals.
In this paper we aim to reveal the spatial resolution
of motion perception and the corresponding receptive
field height of human motion mechanisms through
psychophysical experiments. We make use of an interlea6ed apparent motion sequence consisting of four
frames that alternate between two types of pattern.
Frame-by-frame the pattern shifts successively in one
direction through 90 steps of spatial phase (Fig. 1) and
with each step the pattern switches from one type to the
other (symbolized by the two types of shading in Fig.
1). In general, the two types could be different colours,
different kinds of modulation, different spatial frequencies and so on. The motive for using this sequence is
that it consists of two stationary, counterphase flickering patterns A and B (A: 0/180 phases, B: 90/270),
interleaved in space and time (Shadlen & Carney, 1986;
Georgeson & Shackleton, 1989). Thus any mechanism
that is sensitive only to pattern A or only to pattern B
can sense flicker, but not motion. To detect motion

0042-6989/00/$ - see front matter 2000 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 4 2 - 6 9 8 9 ( 9 9 ) 0 0 2 1 9 - 9

746

M.A. Georgeson, N.E. Scott-Samuel / Vision Research 40 (2000) 745758

Fig. 1. The interleaved apparent motion sequence. Spatial phase steps

through 90 on each frame (two cycles are shown) and the pattern
alternates in some other respect (e.g. between orientations A and B)
with each step. Apparent motion is possible only if there is a
mechanism that can integrate the two different spatial patterns A and
B symbolised by the checkered and shaded boxes, yielding motion
from temporally and spatially interleaved flicker.

there must be a mechanism sensitive to both A and B.

Hence the interleaved sequence can be used as a general
tool for studying the pattern selectivity of motion
mechanisms. It has been used already to demonstrate
that motion detectors cannot combine luminance-modulated and contrast-modulated signals (Ledgeway &
Smith, 1994) at least when distortion products are
absent (Scott-Samuel & Georgeson, 1999).
If patterns A and B contain no common Fourier
components then interleaved motion sequences are
drift-balanced (Chubb & Sperling, 1988) because their
spatial Fourier components are flickering, not moving.
There is no net motion energy in any particular direction. However, they are not microbalanced, because
viewing the images through an aperture may reveal the
presence of local motions in different directions. Receptive fields (RFs) are effectively an aperture through
which motion is computed, and so it may be that by
asking subjects to discriminate the direction of local
motions in these sequences we can obtain estimates of
the RF sizes for motion detection. The logic of this
approach is explained further below.
The experiments also address the orientation selectivity of motion mechanisms, because questions about
orientation selectivity and about RF size are closely
linked: a mechanism with narrow orientation tuning
must have extensive lengthwise integration (a long RF)
while broad orientation tuning implies a short RF. For
a linear filter the RF length and orientation bandwidth
are inversely related. In several experiments we examined the orientation range over which motion sensors
can integrate different grating components. This was
done firstly by alternating the grating orientation at
each step in phase, and then by using modified sequences that progressively overcame some unexpected
difficulties. van den Berg, van de Grind and van Doorn
(1990) tested the orientation selectivity of motion detec-

tion with arrays of randomly oriented line segments or

edges that changed orientation systematically with each
motion step. Rotations of 2540 between steps were
sufficient to eliminate direction discrimination and so
they concluded that the early stage of bilocal motion
detectors is quite narrowly tuned for orientation. As
we shall see, the orientation range for motion perception can be very large up to 980 and this has
the rather different implication that motion detection is
a very local affair, depending on local phase shifts in
the image.
In the third experiment we aim to quantify more
directly the spatial extent over which motion signals are
integrated in the direction orthogonal to the motion.
The display contained spatially interleaved strips of
sinusoidal gratings moving in opposite directions, and
the idea here is that, by analogy with static spatial
acuity, the narrowest strips for which opposite directions can still be resolved will indicate the scale of
spatial integration. A similar logic has been used in
experiments on motion segregation, with random dots
drifting in opposite directions across adjacent strips
(Koenderink, van Doorn & van de Grind, 1985). In
that study the minimal width for resolving the separate
strips of motion was about 8 min arc in foveal vision at
speeds less than 1/s, but rose markedly with both
speed and eccentricity. One limitation of experiments
with random dots is that the dot array is spatially
broadband and so may activate a multiplicity of filters
at different spatial scales under different conditions.
Our experiments used moving sinusoidal waveforms in
order to anchor the scale of analysis to a particular
spatial frequency. Fredericksen, Verstraten and van de
Grind (1997) raised a host of objections to previous
methods of estimating motion RF size. Fortunately,
these potential problems do not apply to our
suprathreshold, acuity-based method, and it therefore
serves as a valuable re-assessment of earlier estimates
(e.g. Gorea, 1985; Anderson & Burr, 1987, 1991).
As noted above, RF height and orientation selectivity
are linked, so it should be possible to derive values for
both with the same set of stimuli. In the context of
psychophysical studies, we adopt the same notion as
Anderson and Burr (1991): The phrase psychophysical
receptive field is used to denote the fact that we are
measuring the psychophysical properties of perceptual
units and not the physiological properties of single
cells.

2. Experiment 1: alternately tilted gratings

Pilot observation of a sequence of sinusoidal gratings
(Fig. 2) tilted alternately to either side of the vertical
and stepping 90 in phase with each presentation
yielded the surprising observation that the display had

M.A. Georgeson, N.E. Scott-Samuel / Vision Research 40 (2000) 745758

747

fairly low contrasts. The mean luminance of the display

was 70.5 cd/m2, and its linearity was calibrated with a
Minolta LS-110 photometer. Contrast was expressed as
Michelson contrast, 100(Lmax Lmin)/(Lmax +Lmin),
where Lmax and Lmin were the maximum and minimum
luminances, respectively.
The interleaved apparent motion sequence consisted
of a sinusoidal grating whose orientation alternated
between + u and u from the vertical with each 90
step in spatial phase (Fig. 2). The orientation u was 0,
15, 30, 45, 60, 75 or 90 from vertical. The temporal
frequency of the sequence was 7.5 Hz and the presentation duration was 266 ms, making each sequence two
temporal cycles (eight frames) long. Contrast was fixed
at 20% and the spatial frequency of the gratings was 1.0
cpd in a 4.5 field (512 512 pixels; 16.8 16.8 cm) at
a viewing distance of 214 cm. On each trial there was a
single presentation, and observers were required to
report both the axis along which they saw motion
(horizontal or vertical; the axis task), and the direction
of motion of the strip that passed over a small, central
fixation point (the direction task). Viewing was binocular, and the observers head was held in position with a
chin and forehead rest. The two authors acted as
observers with 48 trials per observer per condition.

2.2. Results: motion along two axes

Fig. 2. Local phase changes in the tilted gratings. Experiment 1.
Consecutive frames of three motion sequences (a, b, and c); spatial
phase steps by 90 from frame to frame. (a) Vertical grating; (b)
orientation alternates 9 30 from vertical; (c) orientation alternates
960. The numbered strips are stacked to form spacetime diagrams
(below); for 1 3, time increases down the y-axis; for 4 and 5, time
increases from left to right. Strip 1 drifts to the left; strip 2, the
central strip of (b), drifts to the left while the adjacent strip 3 drifts to
the right; strips 4 and 5, central and adjacent in (c), drift upwards and
downwards, respectively.

the appearance of a vertical stack of horizontal strips of

grating, with alternate strips moving in opposite directions (left and right). This percept changed to one of
vertical strips moving up and down when the tilt of the
gratings exceeded 45. Thus consistent local motion was
seen when interleaved flickering gratings were displayed
at different orientations. The following experiment was
designed to investigate this intriguing effect more
rigorously.

2.1. Method
Stimuli were generated by a PC with custom-written
Pascal software, and displayed via a Cambridge Research Systems VSG2/2 8-bit framestore on a gammacorrected Eizo Flexscan 9060S monitor with a refresh
rate of 60Hz. Use of two palette chips together gave the
system the equivalent of 12-bit luminance resolution;
that is, the full 8-bit greyscale was available even at

The results (Fig. 3, filled symbols) confirmed that for

orientations u5 45 strips of horizontal motion were
seen, but at larger angles vertical motion was seen.
Both observers reliably reported the correct direction
(left/right or up/down) of the central strip, although
MAG showed a dip in accuracy at 45 where the axis of
motion was somewhat ambiguous. For NSS the transition between axes occurred sharply between 45 and 60,
and recognition of direction was near 100% for all
angles tested.

2.3. Discussion: local phase shifts predict the percei6ed

motion
An examination of the local phase changes in the
experimental stimuli provides an explanation for the
pattern of results. Fig. 2(ac) shows three different
sequences of the seven used, where u= 0, 30 and 60. In
sequence (a), motion is seen to the left, as spacetime
diagram (1) shows. In sequence (b), two adjacent strips
have opposite direction (2 and 3); the local phase of the
central strip (2) moves leftwards, whilst its neighbour
(3) moves to the right. In sequence (c), the motion axis
has now flipped to vertical, and adjacent strips move
upwards (4) and downwards (5). Despite the differences
in orientation of the component gratings, and the lack
of moving Fourier components, these local phase
changes result in reliable perception of local moving
strips.

748

M.A. Georgeson, N.E. Scott-Samuel / Vision Research 40 (2000) 745758

Fig. 3. Strip direction and axis results. Experiment 1. Results for two subjects (NSS and MAG) for both the axis task (filled circles, solid line)
and direction task (open squares, dashed line), expressed as % correct for the direction task (left hand ordinate), and as % judged horizontal for
the axis task (right hand ordinate). The x-axis shows the orientation u for each sequence.

The dip in direction performance for observer MAG

at u =45 (Fig. 3) indicates that the direction task is
confounded by an uncertainty in the axis task at this
value. A dip was not observed for observer NSS because there was a sharp transition in the axis task
between u= 45 and 60, and so no uncertainty was
present for this observer. This transition from left/right
to up/down does not mean that there is no energy along
the orthogonal axis; for all values 0B u B 90 there is
both horizontal and vertical local motion present, as we
show below. The transition from horizontal to vertical
motion perception presumably reflects the relati6e 6isibility of these two motions.
A grating of spatial frequency f and orientation u has
horizontal and vertical periodicities (u, 6) where u =
f cos u and 6 =f sin u. In the Appendix we show that if
motion perception is based on local phase shifts, then
the sequences used here contain horizontally moving
strips whose height is 1/(4 6), one quarter of the vertical
period (1/6). Informal observation confirmed that this
was indeed the height of the perceived strips. Interestingly, the switching of orientations can also be described as having constant 6 with alternation between u
and u (see Fig. 4) and by an entirely similar argument one can show that there are also vertical strips of
upward and downward phase shift whose width is 1/(4
u).
In principle, then, there are two alternative interpretations for each of our sequences horizontal or
vertical strips of motion. However, when 0 5 u B 45,
6 B u, and the horizontal strips are wider than the
vertical ones, but when 45B u 590 6 \ u and the
vertical strips are wider than the horizontal. A likely
explanation for the crossover from horizontal to vertical motion perception (Fig. 3) is that two mechanisms,
orthogonally tuned, respond to the two axes of motion.
We assume that the two mechanisms have peak sensi-

tivities on the u-axis and 6-axis, respectively (Fig. 4).

Grating components close to the u-axis (u B 45) will
evoke a horizontal motion response more strongly than
vertical, and vice versa for u\ 45. Since there was little
or no evidence for ambiguity or transparency in the
axis task (Fig. 3) it seems likely that the stronger
response suppresses the weaker winner takes all. It
follows that, in spatial terms, the wider strips determine
the perceived axis of motion. One might expect the
crossover point to be at u = 45, but experimentally the
crossover was between 45 and 60. It may be that
extensive practice on horizontally moving stimuli created some bias in favour of the horizontal
interpretation.
We did not observe oblique motion. It might seem
surprising that the vertical and horizontal motions embedded in this animation do not sum to form oblique
movements. Alternate rows (and columns), however,
have opposite directions of movement, and so the pattern of oblique movements obtained by summation
would form a checkerboard in which every small square

Fig. 4. Two orthogonal motion detectors. Experiment 1. (a) u B 45

gives horizontal motion, and (b) u \45 gives vertical motion. The
grey ovals show (a) vertically oriented detectors, (b) horizontally
oriented. Note that, as u varies, the stimuli follow a circular path
through u 6 space, because grating spatial frequency was fixed
[(u 21 +6 21) =(u 22 +6 22) =constant].

M.A. Georgeson, N.E. Scott-Samuel / Vision Research 40 (2000) 745758

749

tion sensitivity of the detectors the stimulus sequence

used here is problematic. Different detectors are recruited at large and small angles, and so there is no way
to test the responsiveness of the horizontal motion
detector when u\ 45. It was therefore necessary to
devise a stimulus sequence that would anchor the motion to the horizontal direction and test only one
detector.

3. Experiment 2: the grating-plaid sequence

Fig. 5. The grating-plaid sequence. Experiment 2. Below: four frames

of the grating-plaid motion sequence (phases 0, 90, 180, 270; alternate frames of grating and plaid). Above: Fourier representation of
the vertical grating (white dots) and the plaid components (black
dots). Both fall within the same vertical detector (light ovals), but
only the plaid components fall within the horizontal detector (dark
ovals).

Fig. 6. Local phase shifts in the grating-plaid sequence. Experiment 2.

Consecutive frames of two motion sequences; spatial phase steps by
90 from frame to frame. (a) Vertical grating; (b) vertical grating
alternates with a plaid. Numbered strips are stacked to produce
space time diagrams in which time is increasing down the y-axis.
Strip 1 drifts to the left, while strips 2 and 3 (adjacent in the
grating-plaid sequence) drift to the right and left, respectively.

had a direction orthogonal to its neighbours. Along the

diagonals of the checkerboard, adjacent squares have
opposite directions which presumably cancel each
other, or lose out in the competition with the coherent
horizontal (or vertical) motion available in rows (or
columns).
This is a reasonably satisfying account of an intriguing motion phenomenon, but for revealing the orienta-

To restrict motion to just one axis, we used the

grating-plaid sequence (Fig. 5). The critical feature of
this sequence is that the alternate frames of vertical
grating prevent the mechanism for vertical motion
(dark shaded ovals) from detecting anything other than
flicker, as this mechanism is sensitive only to the counterphasing plaid components. The mechanism for horizontal motion (light shaded ovals), on the other hand,
is able to detect both the plaid and grating components,
and can therefore produce motion signals by integrating these two images over time. Initial observations
confirmed that this sequence produced horizontal, but
not vertical, strips of apparent motion (see Fig. 6).
From Fig. 5 it can be seen that increasing the angle
(u) of the plaid components from the vertical while
maintaining constant horizontal periodicity (u) has the
effect of increasing the vertical spatial frequency (6).
This means that, as before, the strips of motion are
thinner with increasing u, but now the height of the
strips is 1/(2 6). At sufficiently high values of u, the
plaid components will fall outside the detector (light
grey ovals in Fig. 5) and motion perception will fail.
The limiting value, umax, gives a measure of the orientation range of that detector.
The height of the detectors RF may also be inferred
from the grating-plaid sequence. We assume that when
equal and opposite amounts of motion energy fall
within the receptive field the response of that detector
will be zero. It follows that when two oppositely moving (but otherwise identical) strips exactly cover the
detector, no motion signal will arise (see Fig. 7b). This
is not so if less than two strips fall on the detector (Fig.
7a), in which case there is a predominant motion signal
in one direction, but it remains so with more than two
strips (Fig. 7c). Thus at umax, the height of the strips in
the stimulus sequence should equal half the height of
the motion detector. This strip height is the critical strip
height.

3.1. Method
Apparatus was similar to experiment 1. The interleaved grating-plaid sequence had a temporal frequency
of 7.5 Hz, and was viewed for 266 ms (two cycles). The

750

M.A. Georgeson, N.E. Scott-Samuel / Vision Research 40 (2000) 745758

Both the fixation point and side cursors were displayed

for 250 ms before each trial. The contrast of the single
gratings was fixed at 20% but the plaid contrast varied
across sessions, taking values of 20, 40 or 80%. This
was to compensate for any possible attenuation of the
plaid components due to their off-vertical orientation
and higher spatial frequency [ f=
(u 2 + 6 2)]. Percent
correct direction judgements were plotted against u and
limiting values umax were derived at 75% correct from a
logistic curve fitted to the psychometric function data.

3.2. Results: motion recepti6e fields are 6ery short

Fig. 7. Inferring the receptive field height. White rectangle represents

the receptive field (RF) of a motion detector; light and dark grey bars
are leftward and rightward strips of the grating-plaid sequence. (a)
Most of the RF is occupied by a rightward-moving strip, and the
detector signals rightward motion. (b) Equal areas of the RF are
occupied by rightward- and leftward-moving strips, yielding an ambiguous signal. (c) Similar to (b) but more strips occupy the RF. Thus
when strip height decreases to the critical strip height (b) motion
signalling breaks down. RF height is given by twice the critical strip
height (but see text for exact definition of height H).

stimulus spatial frequencies were changed by adjusting

the viewing distance across sessions; the four viewing
distances (20, 71, 214 and 642 cm) gave horizontal
spatial frequencies of 0.10, 0.33, 1.00 and 3.00 cpd for
both grating and plaid components. The horizontal
periodicity of all stimuli was 4.5 cycles/image. Observers were required to indicate the direction of motion of
a central strip in a single interval, binary choice task.
This centrally located strip varied in height according
to the vertical spatial frequency (6) of the plaid elements of the motion sequence, which varied across
trials within a single session. A central fixation point
was provided and cursors on each side of the display
indicated the location and height of the central strip.

The results for observer NSS (see Fig. 8, left) show

that the critical angle, umax at 75% correct, was greater
than 80 at the lower spatial frequencies, decreasing to
around 60 at the highest spatial frequency tested. This
indicates that motion mechanisms are sensitive over a
very broad range of orientations. Note also that when
u= 80 the radial spatial frequency ( f ) of the plaids
components is 5.8 times higher (2.5 octaves higher)
than the grating, showing that the motion detector can
combine information from components differing very
widely in both orientation and spatial frequency. The
key factor is the coherence of local phase changes. At
all four spatial frequencies (u values) there was no
systematic effect of changing plaid contrast.
In general, 6/u= tan u, and so from the observed
umax values we derive a corresponding 6max which is the
spatial frequency of the strips themselves (not the SF of
the content of the strips) at the point where motion
directions can just be resolved:
6max = u tan umax

(1)

where u and 6 are the vertical and horizontal spatial

frequencies of the plaid, and umax is the critical plaid
angle (see Fig. 5). If we assume that the vertical profile
of the receptive field is Gaussian then its extent can be
represented as 2.5sy, the Gaussian equivalent width,
which is the width of a rectangle that has the same
height and area as a Gaussian with standard deviation

Fig. 8. Orientation range and RF height. Experiment 2. Orientation range in degrees (left) and RF height (H) in cycles (right) plotted against
horizontal spatial frequency for one observer (NSS).

M.A. Georgeson, N.E. Scott-Samuel / Vision Research 40 (2000) 745758

Table 1
Experiment 2: spatial and temporal frequency, speed and performance for NSS expressed in six different waysa
0.1
7.5
75
84
0.95
0.53
0.42
0.08
50.45
a

0.33
7.5
22.5
84
3.14
0.16
0.13
0.08
15.29

1
7.5
7.5
77
4.33
0.12
0.09
0.18
11.08

3
7.5
2.5
63
5.89
0.08
0.07
0.41
8.15

u, cpd
TF, Hz
Speed, deg/s
umax,
6max, cpd
Critical strip, deg
sy, deg
H, cycles
RF height, min arc

Bottom row shows RF height estimated as 2 sy converted to min

Fig. 9. Local phase in the strips stimuli. Experiment 3. Consecutive

frames of two motion sequences. (a) Vertical grating; the four frames
have phases 0, 90, 180 and 270, reading from left to right; (b) is
constructed from two oppositely-drifting gratings with phases 0, 90,
180 and 270, and 270, 180, 90 and 0, which are cut up horizontally
and reassembled as alternate strips. The numbered slices are stacked
to produce space time diagrams in which time is increasing down the
y-axis. Strip 1 drifts to the left, while 2 and 3, adjacent strips in
sequence (b), drift to the left and right, respectively.

sy. The angle subtended by two strips is one vertical

period of the plaid (1/6), and thus sy can be estimated
from the assumption outlined above that the equivalent
height of the Gaussian equals two critical strip heights:
2.5 sy =

6max

(2)

Combining Eqs. (1) and (2) we get:

sy =

4
10 u tan umax

the results, summarized in Table 1, was the shortness of

the receptive fields, which ranged from as little as 0.08
cycles at 0.1 and 0.3 cpd, to 0.4 cycles (8 min arc) at 3.0
cpd.

3.3. Discussion: extraneous cue to motion, and the strip

uncertainty problem

arc.

751

(3)

Anderson and Burr (1991) expressed their RF height

estimates (H) in cycles of the tested spatial frequency,
as H = 2 u sy, and so for direct comparison we have
expressed our results the same way. Receptive field
height (H, in cycles) is plotted against horizontal spatial
frequency (u) in Fig. 8, right. H was not constant, but
increased with spatial frequency. A striking feature of

The second observer (MAG) reported the existence

of an extraneous cue to direction in the grating-plaid
sequence. An image-processing analysis of the gratingplaid sequence revealed that visual temporal integration
over consecutive frames gave rise to an extraneous
static cue to the direction of motion of the central strip.
The cue had the appearance of a stationary, spatially
periodic wiggly distortion across the display in the
vertical direction, with spatial frequency 6. The spatial
phase of this distortion signalled the direction of motion of the central strip. MAGs data were therefore
discarded. The data from NSS, who did not notice this
effect, were apparently unaffected by it.
The experiment was designed to reveal the orientation range and receptive field height of 1st-order motion detectors, but the decline in performance with
decreasing strip height that was used to estimate these
values could have been influenced by uncertainty as to
which strip was to be judged. Uncertainty over which
strip was being marked by the cursors could make
performance fall to chance prematurely. It is therefore
possible that the true receptive field height is even
shorter (and the orientation range wider) than our
estimates. Furthermore, if the filter were polar separable in the u6 domain, rather than Cartesian separable,
then this would also result in an underestimation of the
receptive field height. The results must therefore be
regarded as conservative estimates of these quantities.

4. Experiment 3: the moving strips sequence

To eliminate the artefactual cue described above, a
third sequence was constructed by taking two drifting
gratings, one moving to the left, the other to the right,
and cutting both into horizontal strips of a given
height. These strips were then stacked on top of one
another, alternate strips coming from one or other of
the two original gratings, to form the strips sequence
(Fig. 9).
Comparison of Fig. 9 with Fig. 6 reveals that the
local spacetime diagrams of the grating-plaid and
strips sequences are very similar. In fact, by shifting the
relative phases of the two original gratings, it is possible
to make a strips sequence that is practically indistinguishable from the grating-plaid sequence. In Fig. 9b,
the phases used for the two constituent gratings are

752

M.A. Georgeson, N.E. Scott-Samuel / Vision Research 40 (2000) 745758

0 90180270 and 270 180 90 0. This arrangement ensures that adjacent strips are always 90 out of
phase. An image-processing analysis of the new sequence confirmed that it did not contain a static cue
like that found in the grating-plaid sequence.

in experiment 3, compared with experiment 2. This may

be related to the greater sharpness of the motion
boundaries in the strips sequence, compared with the
grating-plaid sequence.

4.1. Method

5. Experiment 4: controlling for contrast and temporal

frequency

The apparatus and procedure were the same as experiment 2. The only difference was that the alternating
frames of grating and plaid were replaced by consecutive frames of the strips sequence, and that all frames of
the sequence were displayed at 20% contrast. The two
authors and a naive volunteer acted as observers. Critical strip heights (at 75% correct) were estimated from
the fitted psychometric functions, and converted into
RF height estimates (H) as before.

4.2. Results: confirmation of the grating-plaid data

Fig. 10 shows the receptive field height estimates (H,
in cycles) for all three observers, along with the data for
observer NSS from experiment 2 (the grating-plaid
sequence). Although there were slight variations across
observers, the pattern of results was similar. The results
from the strips sequence for NSS were encouragingly
similar to the data from the grating-plaid sequence,
confirming that the extraneous cue in experiment 2 had
not affected NSSs data. As with the grating-plaid
sequence, the most striking feature of the results is the
shortness of the estimated receptive field heights, varying from around 0.1 to 0.4 cycles of the grating period
over the spatial frequency range tested (0.1 3.0 cpd).
The orientation range implied by these receptive field
heights is extremely wide.
RF height estimates for observer NSS showed a
slight but consistent decrease at each spatial frequency

In experiment 2, the grating frames had 20% contrast

while the contrast of the plaid frames was 20, 40 or
80%; the level of plaid contrast had little or no effect
upon direction discrimination performance. In experiment 3 the strips sequence was displayed at 20% contrast only. However, it is possible that performance
might improve with increasing contrast because more
energy is put into the motion mechanism. The critical
strip height might decrease, and the orientation range
would increase. Such contrast dependence would undermine our claim to be measuring RF height. Another
potentially important variable is temporal frequency:
the height estimates in experiments 2 and 3 were obtained at 7.5 Hz drift rate. It might be that larger RFs
are recruited at higher speeds, in which case our estimates would increase with temporal frequency.

5.1. Method
The methods were as experiment 3, except that spatial frequency was 1.0 cpd in all cases, and across
different sessions either (i) the contrast of the strips
sequence was varied (5, 10, 20, 40 or 80%) at fixed
temporal frequency (7.5 Hz), or (ii) the temporal frequency of the stimulus sequence was varied (15.0, 7.5,
3.75 or 1.875 Hz) at fixed contrast (20%). The former
condition was the contrast condition, the latter the
temporal frequency condition. The duration of the display was constant at 266ms in all cases, and in condition (ii) this meant that only half a temporal cycle was
shown at 1.875 Hz. To prevent the motion sequence
from appearing jerky (undersampled), 45 phase shifts
between consecutive frames were used at 1.875 Hz
instead of the 90 phase shifts employed at the other
temporal frequencies. Both authors acted as observers
in the contrast condition (i), NSS in the temporal
frequency condition (ii).

5.2. Results

Fig. 10. Grating-plaid and strips sequences compared. Results for

three observers (AKS, MAG and NSS) for experiment 3 (the strips
sequence), and one observer (NSS) for experiment 2 (the grating-plaid
sequence). Receptive field height H in cycles is plotted against horizontal spatial frequency on loglog axes.

(i) RF height estimates (H, in cycles) for the two

observers (MAG and NSS) in the contrast condition
are shown in Fig. 11(a). At contrasts below 20% H
increased markedly, whereas at and above 20% contrast
H was fairly constant. (ii) There was no systematic
variation in critical strip height across temporal frequency (Fig. 11b), implying that the value of 7.5 Hz
used in the earlier experiments was appropriate.

M.A. Georgeson, N.E. Scott-Samuel / Vision Research 40 (2000) 745758

Fig. 11. The effect of contrast and temporal frequency. (A) Experiment 4(i). Receptive field height H in cycles plotted against the
contrast of the strips sequence for two observers (MAG and NSS) at
a spatial frequency of 1.00 cpd. (B) Experiment 4(ii). Direction
discrimination performance against strip height (in pixels) for NSS at
four speeds (temporal frequencies) at 1 cpd, 20% contrast. 1 pixel =
0.53 min arc. Threshold strip height was about 6 min arc at all
speeds.

753

and 3 (Fig. 10) are representative estimates. Since critical strip height reached its asymptotic level near 20%
contrast, this finding validates the results from experiment 3 carried out at 20% contrast. RF height estimates would not have been much smaller at higher
contrasts. This suggests that a structural limit, imposed by the receptive field height itself, has been
reached at about 2040% contrast. At lower contrasts
reduction of signal:noise ratio may cause performance
to be worse than this structural limit.
Variation of temporal frequency had no systematic
effect on receptive field height estimates, at least up to
15Hz as tested here. Thus spatial resolution for shearing motion is determined by local spatial frequency
(u), not by speed of movement. This conclusion contrasts with the findings of Koenderink et al. (1985)
who found that the strip width required for spatial
segregation of strips of random dots moving in opposite directions increased markedly with speed, being
about 8 min arc at 1/s but 100 min arc at 10/s. The
difference may be largely due to differences in the type
of motion: shearing motion (along the strips) in our
experiments versus expansive/compressive motion
(across the strips) in the Koenderink study. As Koenderink et al. point out, the segregation limit in their
experiment was probably temporal rather than spatial:
the dots had to spend about 70100 ms traversing one
strip in order for the strips to be segregated. In an
analogous study using shearing motion of drifting random dots, Nakayama and Silverman (1984); their Fig.
2) found that strips of motion only 13 min arc high
could just be resolved at 16/s (Nakayama & Silverman, 1984) which is much more in line with our findings.

6. General discussion

Fig. 12. RF height estimates. Averaged results for one subject (NSS)
from experiments 2 and 3 (grating-plaid and strips sequences), compared with data from Anderson and Burr (1991). RF height in cycles
is plotted against horizontal spatial frequency.

5.3. Discussion: pre6ious parameter choices justified

The results from these two control experiments indicate that the values of RF height from experiments 2

Fig. 12 shows a summary of the average RF height

estimates from experiments 2 and 3 for observer NSS,
along with those of Anderson and Burr (1991) who
used a spatial summation technique to measure both
the width and height of motion receptive fields at
threshold. These authors published similar results from
masking experiments (Anderson & Burr, 1987, 1989,
1991; Anderson, Burr & Morrone, 1991), and
confirmed that the RF size (in cycles) increases with
increasing spatial frequency. They also showed that the
heights and widths of the receptive field envelopes were
equal. There is excellent agreement between Anderson
and Burrs (1991) results and our own, despite the very
different methods and theoretical assumptions used. At
all spatial frequencies the estimated receptive field
height is a small fraction of one spatial period. The
heights range from 0.08 cycles at 0.1 cpd to 0.38 cycles
at 3.0 cpd, corresponding to orientation ranges (in the

M.A. Georgeson, N.E. Scott-Samuel / Vision Research 40 (2000) 745758

754

plaid experiment)
respectively.

from

about

9 84

to

9 65,

Fm (u, 6, w)=(2 p i )m u m 1(u+w)F(u, 6, w)

6.1. The Gaussian deri6ati6e model

Do the estimates given above for RF height and
orientation range correspond to plausible space time
filters in motion analysis? To begin to answer this
question we consider a simple, general model for motion filtering. Direction-selective linear filters have receptive fields that are oriented in space time (Adelson
& Bergen, 1985) and it is well known that such fields
can be formed by linear combination of pairs of RFs
that are 90 out of phase with each other in space and
time. This approach gives a good account of directionselective simple cells (McLean & Palmer, 1989) and
complex cells (Emerson, Bergen & Adelson, 1992) in
the primary visual cortex. An analytically tractable
form for this kind of model is based on Gaussian
derivatives (GDs) in space and time (Young, 1985;
Adelson & Bergen, 1986; Johnston, McOwan & Buxton, 1992; Young & Lesperance, 1993). Thus if G(x, y,
t) is a Gaussian function of space and time then Gx and
Gt (where subscripts denote partial differentiation with
respect to a given variable) are suitable subunits that
form an oriented space time receptive field (M1) when
added together: M1 =Gx +Gt. Differentiating again
with respect to x gives us M2 =Gxx +Gxt, which forms
a direction-selective filter whose SF bandwidth is narrower than that of M1. See Bruce, Green and
Georgeson (1996) for illustration (their Fig. 8.7).
Clearly, differentiating m times with respect to x can
define a family of filters Mm (m = 1,2,3,4, ). M1 and
M2 are odd and even filters in space time and can be
used as elements of the Adelson Bergen energy model
to encode velocity exactly. There is an equivalence
between this version of the energy model and at least
one form of the multichannel gradient model (Adelson
& Bergen, 1986; Johnston et al., 1992; Bruce et al.,
1996).
We now derive the spatio-temporal frequency response of Mm in order to see how such filters would
respond to the grating-plaid sequence of experiment 2.
Since motion filters are not space time separable, the
2-D spatial frequency response depends on temporal
frequency, and the equations must be developed in 3-D
(spacetime) not 2-D space alone. Thus we define:
Mm (x, y, t)=

( mG
( mG
+ m1
m
(x
(x
(t

(4)

where G(x, y, t) is a Gaussian function with standard

deviations sx, sy, st. The Fourier transform F(u, 6, w)
of G(x, y, t) is:
F(u, 6, w)= exp[ 2p 2(u 2 s 2x +6 2 s 2y +w 2 s 2t )]

Differentiation with respect to x or t multiplies the

transform by 2 p i uor 2 p i w, respectively. Hence the
Fourier transform (Fm ) of Mm is:

(5)

(6)

Fig. 13 shows iso-sensitivity contours of F2(u, 6, w) in

the (u, 6) plane for w= 7.5 Hz, plotted using the values
of sy from Table 1, and assuming sx = sy, st =0.01 s.
These plots give an idea of the 2-D range of spatial
frequencies that could be combined by the secondderivative based motion filter (Gxx + Gxt ) in order to
extract the motion of strips in the grating-plaid experiment. Positive u values represent motion in the preferred direction, negative u values are the non-preferred
direction. The outermost contour in each plot is the
locus of spatial frequencies (u, 6) for which the filter has
5% of its peak sensitivity at w =7.5 Hz. We note that in
each case the range of plaid frequencies (6max) that
supported perception of moving strips would be nearly
accounted for if the limit of performance was imposed
by the 5% sensitivity contour.
For example, our empirical estimate of sy was 0.09
for a grating of 1 cpd; Fig. 13(C) shows that the M2
filter defined by this spatial scale has a peak sensitivity
at u= 1.9 cpd, and at u= 1 cpd the 5% contour extends
out to 6= 4.0 cpd. This compares well with our experimental 6max of 4.33 cpd (Table 1). Across the four test
spatial frequencies (0.1, 0.33, 1, 3 cpd) the 6max values
from Table 1 are 0.95, 3.1, 4.3, 5.9 cpd, compared with
6= 0.78, 2.5, 4.0, 5.7 cpd at the 5% limit derived from
Fig. 13(AD). The agreement here is fairly good, but
the argument depends on the assumption that human
performance fails when the filters response to one plaid
component falls to about 5% of its response to the
vertical grating (given that the SF of the vertical grating
lies near the peak of the filter). Bearing in mind that
there are two plaid components, the limiting response
(to the plaid) should be about 10% of the response to
the optimal grating. We have independent evidence that
this limit is about right. In a separate study (Georgeson
& Scott-Samuel, 1999) the motion sequence (cf. Fig. 1)
consisted simply of interleaved 1 cpd vertical gratings
of two different contrasts. Recognition of direction fell
to chance when the contrast of one grating fell below
about 10% of the other grating, irrespective of absolute
contrast. This would correspond to a 5% response limit
for each of the plaids components.
In summary, then, the 2-D spatial bandwidth of the
second-derivative motion filter (M2) based upon a circular Gaussian (where sx = sy ) appears well suited to
account for our results on spatial resolution of strips of
motion. The similarity between our RF height estimates
and those of Anderson and Burr (1991) obtained by
very different means lends strong support to this conclusion. Fig. 14 illustrates the short, stubby receptive
fields of the M2 filter that are implied by this analysis.

M.A. Georgeson, N.E. Scott-Samuel / Vision Research 40 (2000) 745758

755

Fig. 13. Gaussian derivative filters for motion. Contour maps of the 2-D spatial frequency response in the (u, 6) plane of the space time oriented
filter Gxx + Gxt (see text) at a temporal frequency w= 7.5 Hz. Solid curve: 95% of peak sensitivity; long dashes: 50%; short dashes: 5%. The 95%
contour exists only in the right half-plane (u\ 0) representing the preferred direction of motion. (A) 0.1 cpd test. Gaussian parameters:
sx =sy = 0.42, st = 0.01 s. (B) 0.33 cpd test; sx = sy = 0.13. (C) 1.0 cpd test; sx =sy =0.09. (D) 3.0 cpd test; sx =sy =0.07. Note: the surface
mapped here does not represent the Fourier transform of a spatial filter kernel, but is the sensitivity in a cross-sectional (u, 6) plane of the
direction-selective spatio-temporal filter that is oriented in the (u, w) plane.

6.2. Orientation tuning

The orientation tuning of the Mm family of GD
filters is easily derived from Eq. (6) by substituting
u =r cos u and 6=r sin u, where r=
(u 2 +6 2) is radial frequency and u is orientation. Assuming sx = sy,
this yields:
Fm (u, r, w)= K(r cos u)m 1(r cos u + w)

width at half-height) increase from 601/m to 601/m 1$,

respectively. Thus orientation tuning of the GD model
filter broadens with increasing temporal frequency, as

(7)

where K =(2 p i )m exp[ 2 p 2(r 2 s 2x +w 2 s 2t )] does not

vary with orientation. Thus the shape of the orientation
tuning curve (Eq. (7)) depends on m, r and w, but not
on spatial scale provided that sx =sy. Orientation tuning curves are usually plotted as a function of u at a
fixed value of spatial and temporal frequency (r, w),
and we see from Eq. (7) that for a static or slowly
moving grating the function (normalized to a peak of
1.0) tends to (cos u)m as w tends to 0, but at high
temporal frequencies if w \ \r then it tends to
(cos u)m 1. The corresponding bandwidths uB (half-

Fig. 14. Subunits of the motion filter at 1 cpd. Spatial receptive fields
of the even (Gxx, left) and odd (Gx, right) subunits of the M2 motion
filter. Gaussian parameters were sx =sy =0.09, as used in Fig.
13(C). Inset, drawn to scale, are strips of a 1 cpd grating at the
critical strip height observed experimentally.

756

M.A. Georgeson, N.E. Scott-Samuel / Vision Research 40 (2000) 745758

observed in adaptation experiments (Snowden, 1992)

and masking experiments (Anderson et al., 1991). The
broadening of model tuning is greatest for m= 1 where
the filter has a half-bandwidth uB of 60 when w =0, but
becomes completely non-oriented when temporal frequency is high. This correctly reflects the behaviour of
adaptation and masking at very low spatial frequencies
(0.1 0.2 cpd) (Kelly & Burbeck, 1987; Snowden, 1992).
When m = 2, uB increases from 45 to 60 with increasing
temporal frequency; when m = 4 it shifts from 32.8 to
37.5, and when m =6 it shifts from 27.0 to 29.5. Thus
increasing the derivative number (m) mirrors the two
effects of adaptation observed by Snowden: with increasing spatial frequency the orientation bandwidths
decrease, and the effect of temporal frequency on bandwidth progressively disappears. Experimental estimates
of spatial frequency bandwidth (in octaves) also become
narrower with increasing peak SF (Wilson, Levi, Maffei, Rovamo & De Valois, 1990) and so an attractive
way to model this important cluster of results in the
Gaussian derivative (GD) framework (Eq. (4)) is to
suppose that for low SF channels m = 1 or 2, but at
higher SFs m increases progressively to around 4 6 (see
Bruce et al., 1996, their Fig. 5.8). Accompanying the
increase in m is a decrease in the spatial scale (sx, sy ).
Many details remain to be worked out, partly because
the available data on bandwidths are fragmentary, and
it also appears that Snowdens (1992) orientation bandwidths are systematically smaller than one would expect
from the GD model with a circular base Gaussian (i.e.
sx = sy ). Physiological data indicate that cells in visual
cortex have quite a range of bandwidths at each centre
frequency (De Valois, Albrecht & Thorell, 1982), and so
it is possible that different psychophysical tasks tap into
different subsets of cells. For example, the tuning revealed in adaptation experiments might reflect the more
narrowly tuned cells at each spatial frequency, while our
experiments may reveal the shortest RFs and hence
more broadly tuned cells. Although Fig. 13 illustrates
the SF tuning surfaces for m =2, our results are consistent with the idea outlined above that the derivative
order m increases with the peak SF of the filter. Setting
m= 3 gave good predicted values of 6max at 3 cpd, but
not at 0.1 or 0.33 cpd where the narrowing of bandwidth (in going from m =2 to m = 3) made the filters
too insensitive at low SFs.

6.3. Off-frequency looking?

Our experiments yield fairly direct estimates of the
RF length factor (sy ), but not the width factor (sx ). In
analyzing the GD model we have supposed that sx = sy,
and shown that this model is consistent with our data.
What happens if sx B Bsy ? This creates a filter with a
more conventional, elongated but narrower RF, and a
higher peak SF with narrower orientation tuning at the

peak SF. However, it turns out that in the 2-D frequency domain (as Fig. 13) the skirts of the M2 filter
response can be quite broad, and that an oriented filter
centred at say 4 or 6 cpd could have sufficiently broad
orientation bandwidth at 6= 1 cpd to be consistent with
our 1 cpd data, though the sensitivity to the grating
component would be low. Such use of an effective, but
less sensitive channel is known as off-frequency looking. It could play a part in our experiments, where
contrast is well above threshold. However, off-frequency looking is unlikely to contribute much to performance at contrast threshold, and so the similarity
between our results and those of Anderson and Burr
(1991) obtained at threshold (Fig. 12) strengthens the
argument that short, stubby receptive fields for motion
do exist in human vision (Fig. 14) (though for a different view, see Watson and Turano (1995)).

6.4. Implications for motion integration in plaids

The very broad orientation selectivity implied by
these short RF heights suggests a surprisingly simple
solution to the aperture problem for coherent plaid
motion. Such broadband detectors would passively integrate plaid components over a wide range of orientations, and hence respond to the movement of local
structure (local phase shifts) in the plaid, as we have
shown. Active schemes for combining the motion components of a plaid pre-suppose that the components are
initially processed separately in narrowband filters. Our
results suggest that this may not be true in general, and
so more explicit schemes for combination such as the
intersection of constraints (IOC) rule (Adelson &
Movshon, 1982) or vectorsum rule (Wilson, Ferrera &
Yo, 1992) might be unnecessary in many cases. Perhaps
we should not be asking how component motions are
combined, but how are they ever segregated? Recent
models (Wilson & Kim, 1994) and experiments (Smith,
Curran & Braddick, 1999) have begun to address this
issue.

Acknowledgements
This work was supported by a BBSRC studentship to
NSS, and by BBSRC grant GR/G63582 to MAG. Parts
of this work have previously been presented in abstract
form (Georgeson & Scott-Samuel, 1994; Scott-Samuel &
Georgeson, 1996).

Appendix A
For experiment 1, the height of the strips seen in the
stimulus sequence can be derived as follows. A sinusoidal grating may be written as:

M.A. Georgeson, N.E. Scott-Samuel / Vision Research 40 (2000) 745758

G(x, y)=sin[2 p(u x +6 y) f]

(A1)

where u and 6 are its horizontal and vertical spatial

frequencies (which together determine its orientation,
u = arctan(6/u)), and f is its phase. The local phase,
f%, expressed relative to an arbitrary position (x1, y1) is
given by:
f%(x1, y1)=f 2 p(u x1 +6 y1)

(A2)

So, for a 090 180 270 sequence with gratings

tilting alternately one way and then the other (where
the horizontal spatial frequency u is constant and the
vertical spatial frequency alternates between + 6 and
6) the local phases in each of the four frames are
given by:
f%1 = 2 p(u x1 + 6 y1)

(A3)

p
f%2 = 2 p(u x1 6 y1)
2

(A4)

f%3 =p 2 p(u x1 +6 y1)

(A5)

f%4 =

3p
2 p(u x1 6 y1)
2

(A6)

We propose that the local direction of perceived

motion depends on the direction of local phase shift.
Thus motion to the right occurs when:
f%1 Bf%2 B f%3 Bf%4

(A7)

and the reverse is true for motion to the left. To

determine the height and direction of the strips
perceived in the sequence, it is necessary to find the
range of y values over which there is a common
direction of motion, i.e. within which f %1 Bf %2 and
f %2 Bf %3. From Eqs. (A3), (A4) and (A5) we require:
p
2 p(u x1 +6 y1)B 2 p(u x1 6 y1)
2

and

p
p 2 p(u x1 + 6 y1) \ 2 p(u x1 6 y1)
2

(A8)

which simplifies to:

y1 \

1
86

and

y1 B

1
86

(A9)

Thus the range of y values that share a common

direction of phase shift is 9 1/(8 6), and the height of
such moving strips is 1/(4 6), one quarter of a period of
6.

References
Adelson, E. H., & Bergen, J. R. (1985). Spatiotemporal energy
models for the perception of motion. Journal of the Optical
Society of America A, 2, 284299.
Adelson, E. H., & Bergen, J. R. (1986). The extraction of spatio-

757

temporal energy in human and machine vision. IEEE workshop

on motion: representation and analysis, Charleston, S.C.
Adelson, E. H., & Movshon, J. A. (1982). Phenomenal coherence
of moving visual patterns. Nature, 300, 523 525.
Anderson, S. J., & Burr, D. C. (1987). Receptive field size of
human motion detection units. Vision Research, 27, 621635.
Anderson, S. J., & Burr, D. C. (1989). Receptive field properties of
human motion detector units inferred from spatial frequency
masking. Vision Research, 29, 1343 1358.
Anderson, S. J., & Burr, D. C. (1991). Spatial summation properties of directionally selective mechanisms in human vision. Journal Of the Optical Society Of America A, 8, 1330 1339.
Anderson, S. J., Burr, D. C., & Morrone, M. C. (1991). Two-dimensional spatial and spatial-frequency selectivity of motionsensitive mechanisms in human vision. Journal of the Optical
Society of America A, 8, 1340 1351.
Braddick, O. J. (1993). Segmentation versus integration in visual
motion processing. Trends in Neurosciences, 16, 263 268.
Bruce, V., Green, P. R., & Georgeson, M. A. (1996). Visual perception: physiology, psychology and ecology (3rd ed.). Hove, UK:
Psychology Press.
Chubb, C., & Sperling, G. (1988). Drift-balanced random stimuli: a
general basis for studying non-Fourier motion perception. Journal of the Optical Society of America A, 5, 1986 2007.
De Valois, R. L., Albrecht, D. G., & Thorell, L. G. (1982). Spatial
frequency selectivity of cells in macaque visual cortex. Vision
Research, 22, 545 559.
Emerson, R. C., Bergen, J. R., & Adelson, E. H. (1992). Directionally selective complex cells and the computation of motion energy in cat visual cortex. Vision Research, 32, 203 218.
Fredericksen, R. E., Verstraten, F. A. J., & van de Grind, W. A.
(1997). Pitfalls in estimating motion detector receptive field geometry. Vision Research, 37, 99 119.
Georgeson, M. A., & Scott-Samuel, N. E. (1994). Orientation
bandwidth of motion perception. Ophthalmic and Physiological
Optics, 14, 438 439.
Georgeson, M. A. & Scott-Samuel, N. E., (1999). Motion contrast:
a new metric for direction discrimination. Vision Research, 39,
4393 4402.
Georgeson, M. A., & Shackleton, T. M. (1989). Monocular motion
sensing, binocular motion perception. Vision Research, 29,
1511 1523.
Gorea, A. (1985). Spatial integration characteristics in motion detection and direction identification. Spatial Vision, 1, 85102.
Johnston, A., McOwan, P. W., & Buxton, H. (1992). A computational model of the analysis of some first-order and second-order motion patterns by simple and complex cells. Proceedings of
the Royal Society B, 250, 297 306.
Kelly, D. H., & Burbeck, C. A. (1987). Further evidence for a
broadband, isotropic mechanism sensitive to high-velocity stimuli. Vision Research, 27, 1527 1537.
Koenderink, J. J., van Doorn, A. J., & van de Grind, W. A.
(1985). Spatial and temporal parameters of motion detection in
the peripheral visual field. Journal of the Optical Society of
America A, 2, 252 259.
Ledgeway, T., & Smith, A. T. (1994). Evidence for separate motion-detecting mechanisms for first- and second-order motion in
human vision. Vision Research, 34, 2727 2740.
McLean, J., & Palmer, L. A. (1989). Contribution of linear spatiotemporal receptive field structure to velocity selectivity of
simple cells in area 17 of the cat. Vision Research, 29, 675679.
Nakayama, K., & Silverman, G. H. (1984). Temporal and spatial
characteristics of the upper displacment limit for motion in random dots. Vision Research, 24, 293 299.

758

M.A. Georgeson, N.E. Scott-Samuel / Vision Research 40 (2000) 745758

Scott-Samuel, N. E., & Georgeson, M. A. (1996). Receptive-field length

of motion sensors in human vision. Perception (suppl.), 25, 125.
Scott-Samuel, N. E., & Georgeson, M. A. (1999). Does early non-linearity account for second-order motion? Vision Research, 39,
2853 2865.
Shadlen, M., & Carney, T. (1986). Mechanisms of motion perception
revealed by a new cyclopean illusion. Science, 232, 9597.
Smith, A. T., Curran, W., & Braddick, O. J. (1999). What motion
distributions yield global transparency and spatial segmentation?
Vision Research, 39, 11211132.
Snowden, R. J. (1992). Orientation bandwidth: the effect of spatial and
temporal frequency. Vision Research, 32, 19651974.
van den Berg, A. V., van de Grind, W. A., & van Doorn, A. J. (1990).
Motion detection in the presence of local orientation changes.
Journal of the Optical Society of America A, 7, 933939.
Watson, A. B., & Turano, K. (1995). The optimal motion stimulus.
Vision Research, 35, 325336.

Wilson, H. R., Ferrera, V. P., & Yo, C. (1992). A psychophysically

motivated model for two-dimensional motion perception. Visual
Neuroscience, 9, 79 97.
Wilson, H. R., & Kim, J. (1994). A model for motion coherence
and transparency. Visual Neuroscience, 11, 1205 1220.
Wilson, H. R., Levi, D. M., Maffei, L., Rovamo, J., & De Valois,
R. L. (1990). The perception of form: retina to striate cortex.
In L. Spillman, & J. S. Werner, Visual perception: the neurophysiological foundations (pp. 231 272). London: Academic
Press.
Young, R. A. (1985). The Gaussian derivative theory of spatial
vision: analysis of cortical cell receptive field line-weighting
profiles. Warren, MI: General Motors Research Labs. GMR4920.
Young, R. A., & Lesperance, R. M. (1993). A physiological model
of motion analysis for machine vision. Warren, MI: General
Motors Research Labs. GMR-7878.