5D Library › Computational Aesthetics, Visual Saliency Maps and Cognitive Ergonomics in User Interface Design · 中文
Chapter 4 How Gaze Moves: Centre Bias, Saccades and Inhibition of Return
A saliency map is only a static map of “what stands out”. Real viewing is a sequence of discrete events: the eye rests in one place (a fixation) for two or three hundred milliseconds, then jumps at great speed to another (a saccade), and rests again. This chapter takes up three questions. How much of the distribution of fixations can low-level saliency explain? What is the temporal structure of fixations? And what role does inhibition of return play in a scanpath?
Basic numbers for fixations and saccades
Rayner’s Bartlett lecture paper (2009) summarised eye-movement parameters in reading, scene perception and visual search (Rayner 2009): in silent reading the mean fixation lasts about 225–250 ms, with saccades of about 2 degrees (7–9 letters); in oral reading fixations last about 275–325 ms; in scene perception fixations last about 260–330 ms with saccades of about 4–5 degrees; in visual search fixations last about 180–275 ms with saccades of about 3 degrees. He stressed that the values for scene perception and visual search depend heavily on the exact task. The framework’s “fixations last about 200–300 ms” broadly falls within these ranges but should be written as “task-dependent, commonly about 180–330 ms”.
Saccade amplitude, duration and peak velocity are related in a stable way that Bahill, Clark and Stark (1975) called the “main sequence” (Bahill et al. 1975): up to about 15–20 degrees, peak velocity rises roughly linearly with amplitude, and larger saccades approach saturation; duration also rises with amplitude, the small saccades of reading lasting only 20–30 ms. The program in this chapter uses a common linear approximation, “duration = 21 + 2.2 × amplitude (degrees) ms”; this is an assumption, not original data checked for this book.
Centre bias and task: how much can saliency explain?
Tatler (2007) found that observers looking at pictures of scenes show a strong “central fixation bias”: fixations cluster at the centre of the image, and the bias is largely independent of the distribution of image features and of the viewing task (Tatler 2007). Web pages add a top-left bias and the influence of reading order. Nielsen (2006), from the eye-tracking records of 232 users, described an “F-shaped” pattern of viewing: a horizontal sweep across the top of the content, a second, shorter horizontal sweep a little lower, and then a vertical scan down the left side (Nielsen 2006). The original article itself says this is a dominant tendency, not a universal rule: on a product page a large image pushes the second bar lower and users concentrate on the price and “add to cart” area; on a search-results page the second bar can be longer than the first. The F-pattern should therefore be read as a conditional description of scanning text-heavy pages.
To show how much low-level saliency can still explain when these factors combine, the program below is the author’s model: 40 synthetic observers each make 12 fixations on the layout-A page of Chapter 3, each fixation drawn from a mixture of four maps—low-level saliency (weight 0.25), centre bias (a Gaussian peaked at the page centre, weight 0.25), a task map (the task is “find the price”, which sits in the body-text area, weight 0.35) and a top-left reading bias (weight 0.15). Several predictor maps are then scored with two common measures: NSS (mean normalised saliency at fixated points, 0 for chance) and AUC (the probability that a fixated location’s predicted value exceeds that of a random location, 0.5 for chance).
Calculation
# How much of a fixation map can bottom-up saliency explain? Synthetic observers
# (the author's model) on the 16 x 12 saliency map of p06, layout A. Each observer's
# fixation density is a mixture: w_s * saliency + w_c * centre bias (Gaussian,
# sd = 1/4 of the page; Tatler 2007) + w_t * task map (here: "find the price", which
# lives in the body-text block) + w_l * top-left reading bias. 40 observers x 12
# fixations are sampled from the true mixture (w_s, w_c, w_t, w_l) = (.25, .25, .35, .15).
# Scores: NSS (mean normalised saliency at fixations) and AUC (fixated vs all cells).
import math, random, statistics, importlib.util, io, contextlib, pathlib, sys
sys.dont_write_bytecode = True
spec = importlib.util.spec_from_file_location("p06", pathlib.Path(__file__).with_name(
"p06_itti_koch.py"))
p06 = importlib.util.module_from_spec(spec)
with contextlib.redirect_stdout(io.StringIO()):
spec.loader.exec_module(p06)
SAL, _ = p06.saliency(p06.page((.85, .1, .1)))
CELLS = [(x, y) for y in range(12) for x in range(16)]
def norm(m):
t = sum(m[y][x] for x, y in CELLS)
return [[m[y][x] / t for x in range(16)] for y in range(12)]
CENTRE = norm([[math.exp(-((x - 7.5) / 4) ** 2 / 2 - ((y - 5.5) / 3) ** 2 / 2)
for x in range(16)] for y in range(12)])
TASK = norm([[1.0 if p06.label(x, y) == "body text" else 0.02 for x in range(16)]
for y in range(12)])
TOPLEFT = norm([[math.exp(-(x + y) / 4) for x in range(16)] for y in range(12)])
SALN = norm(SAL)
def mix(w):
maps = (SALN, CENTRE, TASK, TOPLEFT)
return [[sum(wi * m[y][x] for wi, m in zip(w, maps)) for x in range(16)]
for y in range(12)]
def sample(rng, m, k):
cum, t = [], 0.0
for x, y in CELLS:
t += m[y][x]; cum.append(t)
out = []
for _ in range(k):
u = rng.random() * t
out.append(CELLS[next(i for i, c in enumerate(cum) if c >= u)])
return out
def nss(m, fix):
v = [m[y][x] for x, y in CELLS]; mu = statistics.fmean(v)
sd = math.sqrt(math.fsum((a - mu) ** 2 for a in v) / len(v))
return statistics.fmean((m[y][x] - mu) / sd for x, y in fix)
def auc(m, fix):
allv = [m[y][x] for x, y in CELLS]; pos = [m[y][x] for x, y in fix]
return statistics.fmean((sum(p > a for a in allv) + 0.5 * sum(p == a for a in allv))
/ len(allv) for p in pos)
if __name__ == "__main__":
rng = random.Random(2007); TRUE = (.25, .25, .35, .15)
fix = [f for _ in range(40) for f in sample(rng, mix(TRUE), 12)]
print(f"{len(fix)} fixations from 40 synthetic observers (true mixture {TRUE})")
print(" predictor NSS AUC")
for lab, w in (("bottom-up saliency only", (1, 0, 0, 0)),
("centre bias only", (0, 1, 0, 0)),
("saliency + centre", (.5, .5, 0, 0)),
("task map only", (0, 0, 1, 0)), ("true mixture", TRUE)):
m = mix(w)
print(f" {lab:32s} {nss(m, fix):5.2f} {auc(m, fix):5.3f}")
print(f" {'uniform map (chance)':32s} {0:5.2f} {0.5:5.3f}")
for el in ("body text", "CTA button"):
cells = sum(p06.label(x, y) == el for x, y in CELLS)
share = statistics.fmean(p06.label(x, y) == el for x, y in fix)
print(f" {el:10s}: {share:.3f} of fixations, {cells} cells ="
f" {cells / len(CELLS):.3f} of the page")
# 480 fixations from 40 synthetic observers (true mixture (0.25, 0.25, 0.35, 0.15))
# predictor NSS AUC
# bottom-up saliency only 0.23 0.572
# centre bias only 0.19 0.564
# saliency + centre 0.25 0.587
# task map only 1.02 0.595
# true mixture 1.06 0.674
# uniform map (chance) 0.00 0.500
# body text : 0.227 of fixations, 7 cells = 0.036 of the page
# CTA button: 0.033 of fixations, 4 cells = 0.021 of the pageOver 480 fixations, low-level saliency alone scores NSS 0.23 and AUC 0.572; centre bias alone scores 0.19 and 0.564, barely different; an equal mixture of the two scores 0.25 and 0.587. A predictor that knows only the task (the price is in the body text) scores NSS 1.02 and AUC 0.595—by NSS more than four times low-level saliency. Even the true mixture weights reach only NSS 1.06 and AUC 0.674, because each fixation is random and any predictor has a ceiling. The CTA button receives 3.3 % of fixations while occupying 2.1 % of the page (4 cells): saliency does pull some fixations onto the button. The task area (the body text, 7 cells) occupies only 3.6 % of the page yet receives 22.7 % of fixations, more than six times its share of the area and far more than the CTA. This follows from the model’s set-up: the task map has a weight of 0.35 in the mixture, so the task area receives a large share of fixations, not the great majority.
This is of course an example constructed from assumptions; the mixture weights were set by the author, not measured. Its purpose is to show a structural fact: when fixations are driven jointly by several sources, predicting them from only one source gives mediocre scores, and a content-free baseline such as centre bias may do as well as a carefully computed saliency map. This is consistent with the observation in Borji and Itti’s review (Borji and Itti 2013).
Inhibition of return and scanpaths
Posner and Cohen (1984) found that when attention is drawn to a location by a peripheral cue, detection of targets there is faster for a short time after the cue (facilitation); but after about 300 ms, detection at the same location becomes slower (Posner and Cohen 1984); later studies show that this inhibition can last several seconds (Klein 2000). Klein’s 2000 review interprets this “inhibition of return” as a mechanism that aids search: it makes attention less likely to return to locations just inspected, favouring new ones (Klein 2000). Inhibition of return in the Itti–Koch model is a computational version of this idea; without it the winner-take-all network would stay forever on the most salient location.
The program below turns the saliency maps of Chapter 3 into a timed scanpath (the author’s model): the page spans 30 degrees of visual angle; fixation durations follow a gamma distribution with mean 250 ms and standard deviation 100 ms; saccade durations follow the main-sequence approximation above; each shift goes to the most salient location not currently inhibited; a disk of radius 2 cells around each fixated location is inhibited, and the inhibition lifts after ms ( infinite means permanent inhibition, none at all). So that observers differ, each sees the saliency map multiplied by a random factor ( standard normal, independent across locations, fixed for each observer).
Calculation
# From a saliency map to a timed scanpath (the author's model). Saliency maps of p06
# (layouts A and B). The page is shown 30 deg wide (16 cells: 1.875 deg per cell).
# Each shift goes to the most salient cell not under inhibition (IOR radius 2 cells);
# inhibition recovers after a fixed time tau (tau = infinity: permanent IOR; 0: none).
# Fixation durations: gamma, mean 250 ms, sd 100 ms (within the 180-330 ms task ranges
# summarised by Rayner 2009); saccade duration 21 + 2.2 A ms for amplitude A in deg
# (a common linear main-sequence approximation; an assumption here). Viewers differ:
# each sees the map times (1 + 0.25 z) per cell, z standard normal, fixed per viewer.
import random, statistics, importlib.util, io, contextlib, pathlib, sys
sys.dont_write_bytecode = True
spec = importlib.util.spec_from_file_location("p06", pathlib.Path(__file__).with_name(
"p06_itti_koch.py"))
p06 = importlib.util.module_from_spec(spec)
with contextlib.redirect_stdout(io.StringIO()):
spec.loader.exec_module(p06)
DEG, R = 30 / 16, 2
CELLS = [(i, j) for j in range(12) for i in range(16)]
NEAR = {c: [d for d in CELLS if (c[0] - d[0]) ** 2 + (c[1] - d[1]) ** 2 <= R * R]
for c in CELLS}
LAB = {c: p06.label(*c) for c in CELLS}
def run(S, rng, tau, horizon=3000.0, start=(7, 5), noise=0.25):
S = {(i, j): S[j][i] * (1 + noise * rng.gauss(0, 1)) for i, j in CELLS}
t, (x, y), until, path = 0.0, start, {}, []
while t < horizon:
t += rng.gammavariate(6.25, 40.0) # fixation: mean 250, sd 100
for d in NEAR[(x, y)]:
until[d] = t + tau # inhibited until this time
free = [c for c in CELLS if until.get(c, -1.0) <= t]
if not free:
break
i, j = max(free, key=S.get)
amp = DEG * ((i - x) ** 2 + (j - y) ** 2) ** 0.5
t += 21 + 2.2 * amp; x, y = i, j; path.append((t, LAB[(x, y)], amp))
return path
if __name__ == "__main__":
maps = {k: p06.saliency(p06.page(c))[0] for k, c in (("A", (.85, .1, .1)),
("B", (.35, .35, .35)))}
for k in ("A", "B"):
print(f"layout {k}, noise-free viewer: first 6 fixations"
" (arrival ms, element, saccade deg)")
p = run(maps[k], random.Random(1975), float("inf"), noise=0.0)
print(" " + "; ".join(f"{t:.0f} {lab} {a:.1f}" for t, lab, a in p[:3]))
print(" " + "; ".join(f"{t:.0f} {lab} {a:.1f}" for t, lab, a in p[3:6]))
print("time to first fixation on the CTA, 500 simulated viewers, 3 s horizon")
print(" IOR recovery tau layout A: median found layout B: median found")
for tau in (float("inf"), 900.0, 300.0, 0.0):
row = []
for k in ("A", "B"):
hits = []
for v in range(500):
p = run(maps[k], random.Random(v), tau)
hits.append(next((t for t, lab, _ in p if lab == "CTA button"), None))
ok = [h for h in hits if h is not None]
med = f"{statistics.median(ok):6.0f}" if ok else " -"
row.append(f"{med} ms {len(ok) / 500:6.1%}")
lab = "permanent" if tau == float("inf") else f"{tau:.0f} ms"
print(f" {lab:16s} {row[0]:>21s} {row[1]:>21s}")
stuck = run(maps["B"], random.Random(3), 0.0)
n = len(set(lab for _, lab, _ in stuck))
print(f" without IOR (tau 0) a layout-B viewer stays put: {len(stuck)} fixations,"
f" {n} element(s)")
# layout A, noise-free viewer: first 6 fixations (arrival ms, element, saccade deg)
# 247 CTA button 7.5; 484 logo 21.0; 815 hero photo 9.4
# 1163 hero photo 4.2; 1469 white space 10.1; 1637 hero photo 15.0
# layout B, noise-free viewer: first 6 fixations (arrival ms, element, saccade deg)
# 262 logo 14.6; 475 hero photo 9.6; 794 hero photo 4.2
# 1146 hero photo 5.9; 1455 hero photo 11.4; 1608 white space 8.4
# time to first fixation on the CTA, 500 simulated viewers, 3 s horizon
# IOR recovery tau layout A: median found layout B: median found
# permanent 292 ms 94.8% 2980 ms 5.0%
# 900 ms 300 ms 99.8% 1970 ms 3.6%
# 300 ms 298 ms 99.0% 1118 ms 0.4%
# 0 ms 279 ms 95.6% - ms 0.0%
# without IOR (tau 0) a layout-B viewer stays put: 9 fixations, 1 element(s)A noise-free observer on layout A fixates the CTA first (arriving at 247 ms after a 7.5-degree saccade), then makes a 21.0-degree saccade to the logo and returns to the hero photograph; on layout B the first six fixations are the logo and the hero photograph, and the CTA is not among them. Among 500 noisy observers, on layout A between 94.8 % and 99.8 % fixate the CTA within 3 s, with median arrival times of about 280–300 ms, that is, on the first or second fixation; on layout B only 0.4 % to 5.0 % do, with medians between 1 and 3 s. The duration of inhibition of return hardly matters for layout A, because the CTA is the most salient location from the start; but on layout B, the shorter the inhibition, the more easily attention is pulled back to the photograph and logo, and the share finding the CTA falls from 5.0 % with permanent inhibition to 0.4 % at ms and to 0 with no inhibition. The last line shows the extreme case without inhibition of return: one layout-B observer makes 9 fixations, all on the same element.
Note
The gap between the model and real eye movements. Every step of this model is a simplification. Real fixation locations are influenced by task, text, reading habits and centre bias (previous section); saccades have landing errors and often fall short of the target; fixation durations depend on the processing of the current location rather than being independent random numbers; and inhibition of return is weaker in real viewing than here, since people do look back. The model is useful for turning “more salient elements are attended earlier” into testable numbers and for showing how sensitive that prediction is to the parameters of inhibition of return, not for predicting fixation times on any real page.
Hypothesis · not established · Time to first fixation and time to click
The author’s hypothesis (not established): in “find and click” tasks, the time to first fixate the target button accounts for more than half of the differences in completion time; that is, if the optimised layouts in the Chapter 8 trial shorten completion time, most of the shortening comes from fixating the target earlier rather than from deciding or pointing. What result would show it wrong: in a sub-sample with eye-tracking, the upper confidence limit of the proportion of the total effect mediated by time to first fixation is below 50 %.
Chapter summary
Fixation durations depend on the task, commonly about 180–330 ms (Rayner 2009), and the framework’s “200–300 ms” should be qualified accordingly; saccade amplitude, duration and velocity follow the main sequence. On web pages, centre bias, the top-left and reading order, the task and text all influence fixations; the F-pattern is a conditional description of scanning text-heavy pages. The author’s mixture model shows that when fixations are driven by several sources together, predicting them from low-level saliency alone gives mediocre scores, and a content-free centre bias can do about as well. Inhibition of return moves attention away from locations already inspected; in the simulation it hardly affects a button that is the most salient element anyway, but for a button that is not salient, the shorter the inhibition, the harder it is to find.
References
Fan, Y. (2026). Computational Aesthetics, Visual Saliency Maps and Cognitive Ergonomics in User Interface Design: From the Hick–Hyman Law and Spatial Frequency Channels to Saccadic Trajectory Modelling, Read Beside the Kaogongji and Classical Treatise on Architectural Methods [计算美学、视觉显著性图谱与人机界面认知工程] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23836258 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).