Orientation-specific learning of the prior assumption for 3D slant perception

Shuichiro Taya, Masayuki Sato

Research output: Contribution to journalArticle

Abstract

We usually interpret a trapezoidal image on our retina as a slanted rectangle rather than a frontoparallel trapezoid, because we use a statistical assumption (i.e. rectangles are more common than trapezoids), called a ‘prior’, for recovering the 3D world from ambiguous 2D images. Here we report that the shape prior for recovering 3D slant can be updated differently depending on the slant axis orientation (horizontal vs. vertical). The participants were exposed to a variety of trapezoidal images surrounded by a stereoscopic reference plane. The perspective transformation of the images was interpreted as 2D shape, rather than 3D slant because the surrounding plane enhanced disparity. We found that, after continuous exposure to such images, the participants relied less on the shape information for recovering 3D slant, suggesting the update of priors via experience (i.e., rectangles are less common than trapezoids). Importantly, the learning effect was context (slant-axis) specific although partially transferred across contexts; the training with horizontal-axis slant reduced the reliance on perspective even in vertical-axis slant estimation but not vice versa. The results suggest that context-specific training is vital to update the prior for the horizontal-axis slant, whereas it is not required to update the prior for a vertical-axis slant.

Original languageEnglish
Article number11063
JournalScientific Reports
Volume8
Issue number1
DOIs
Publication statusPublished - 2018 Dec 1

Fingerprint

learning

ASJC Scopus subject areas

  • General

Cite this

Orientation-specific learning of the prior assumption for 3D slant perception. / Taya, Shuichiro; Sato, Masayuki.

In: Scientific Reports, Vol. 8, No. 1, 11063, 01.12.2018.

Research output: Contribution to journalArticle

@article{13c73691c3df434a8e1a59b81229abd6,
title = "Orientation-specific learning of the prior assumption for 3D slant perception",
abstract = "We usually interpret a trapezoidal image on our retina as a slanted rectangle rather than a frontoparallel trapezoid, because we use a statistical assumption (i.e. rectangles are more common than trapezoids), called a ‘prior’, for recovering the 3D world from ambiguous 2D images. Here we report that the shape prior for recovering 3D slant can be updated differently depending on the slant axis orientation (horizontal vs. vertical). The participants were exposed to a variety of trapezoidal images surrounded by a stereoscopic reference plane. The perspective transformation of the images was interpreted as 2D shape, rather than 3D slant because the surrounding plane enhanced disparity. We found that, after continuous exposure to such images, the participants relied less on the shape information for recovering 3D slant, suggesting the update of priors via experience (i.e., rectangles are less common than trapezoids). Importantly, the learning effect was context (slant-axis) specific although partially transferred across contexts; the training with horizontal-axis slant reduced the reliance on perspective even in vertical-axis slant estimation but not vice versa. The results suggest that context-specific training is vital to update the prior for the horizontal-axis slant, whereas it is not required to update the prior for a vertical-axis slant.",
author = "Shuichiro Taya and Masayuki Sato",
year = "2018",
month = "12",
day = "1",
doi = "10.1038/s41598-018-29361-2",
language = "English",
volume = "8",
journal = "Scientific Reports",
issn = "2045-2322",
publisher = "Nature Publishing Group",
number = "1",

}

TY - JOUR

T1 - Orientation-specific learning of the prior assumption for 3D slant perception

AU - Taya, Shuichiro

AU - Sato, Masayuki

PY - 2018/12/1

Y1 - 2018/12/1

N2 - We usually interpret a trapezoidal image on our retina as a slanted rectangle rather than a frontoparallel trapezoid, because we use a statistical assumption (i.e. rectangles are more common than trapezoids), called a ‘prior’, for recovering the 3D world from ambiguous 2D images. Here we report that the shape prior for recovering 3D slant can be updated differently depending on the slant axis orientation (horizontal vs. vertical). The participants were exposed to a variety of trapezoidal images surrounded by a stereoscopic reference plane. The perspective transformation of the images was interpreted as 2D shape, rather than 3D slant because the surrounding plane enhanced disparity. We found that, after continuous exposure to such images, the participants relied less on the shape information for recovering 3D slant, suggesting the update of priors via experience (i.e., rectangles are less common than trapezoids). Importantly, the learning effect was context (slant-axis) specific although partially transferred across contexts; the training with horizontal-axis slant reduced the reliance on perspective even in vertical-axis slant estimation but not vice versa. The results suggest that context-specific training is vital to update the prior for the horizontal-axis slant, whereas it is not required to update the prior for a vertical-axis slant.

AB - We usually interpret a trapezoidal image on our retina as a slanted rectangle rather than a frontoparallel trapezoid, because we use a statistical assumption (i.e. rectangles are more common than trapezoids), called a ‘prior’, for recovering the 3D world from ambiguous 2D images. Here we report that the shape prior for recovering 3D slant can be updated differently depending on the slant axis orientation (horizontal vs. vertical). The participants were exposed to a variety of trapezoidal images surrounded by a stereoscopic reference plane. The perspective transformation of the images was interpreted as 2D shape, rather than 3D slant because the surrounding plane enhanced disparity. We found that, after continuous exposure to such images, the participants relied less on the shape information for recovering 3D slant, suggesting the update of priors via experience (i.e., rectangles are less common than trapezoids). Importantly, the learning effect was context (slant-axis) specific although partially transferred across contexts; the training with horizontal-axis slant reduced the reliance on perspective even in vertical-axis slant estimation but not vice versa. The results suggest that context-specific training is vital to update the prior for the horizontal-axis slant, whereas it is not required to update the prior for a vertical-axis slant.

UR - http://www.scopus.com/inward/record.url?scp=85050592902&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85050592902&partnerID=8YFLogxK

U2 - 10.1038/s41598-018-29361-2

DO - 10.1038/s41598-018-29361-2

M3 - Article

AN - SCOPUS:85050592902

VL - 8

JO - Scientific Reports

JF - Scientific Reports

SN - 2045-2322

IS - 1

M1 - 11063

ER -