Noise causes slant underestimation in stereo and motion

Area of Science:

• Computer Vision
• Computational Neuroscience
• Perception

Background:

• Estimating 3D shape from multiple views is crucial for computer vision and understanding visual perception.
• Image noise introduces bias in 3D shape (surface normal) estimation, leading to significant parameter estimation errors.
• This bias is known to cause underestimation of slant in both computational models and psychophysical experiments.

Purpose of the Study:

• To analyze the inherent bias in 3D shape estimation from motion and stereo vision using orientation disparity.
• To investigate how this bias affects the perception of slant, particularly the anisotropy between horizontal and vertical slant in stereo vision.
• To demonstrate the bias in 3D motion perception through a novel illusory display and discuss optimal strategies to mitigate it.

Main Methods:

• Analysis of orientation disparity in stereo vision to quantify slant underestimation.
• Development of an illusory display to demonstrate bias in 3D motion perception.
• Examination of statistically optimal strategies for 3D shape estimation.

Main Results:

• The inherent bias in 3D shape estimation predicts the underestimation of slant.
• In stereo vision, the bias explains the observed anisotropy in the perception of horizontal versus vertical slant.
• The study successfully demonstrates the bias in 3D motion perception using a novel display.

Conclusions:

• Image noise-induced bias is a significant factor in 3D shape estimation from multiple views.
• Understanding and mitigating this bias is essential for accurate 3D reconstruction and visual system function.
• Statistically optimal strategies and potential adaptations in visual systems are discussed for managing this estimation bias.