Estimating 3D tilt from local image cues in natural scenes

Area of Science:

• Computer Vision
• Computational Neuroscience
• 3D Perception

Background:

• Estimating 3D surface orientation is crucial for 3D shape perception.
• Local image cues offer potential information for inferring surface orientation.

Purpose of the Study:

• To investigate how disparity gradient, luminance gradient, and texture orientation cues combine for 3D tilt estimation in natural scenes.
• To determine the Bayes optimal combination of these cues for tilt estimation.

Main Methods:

• Collected a database of natural stereoscopic images with ground-truth range data.
• Analyzed the relationship between ground-truth tilt and local image cue values.
• Calculated Bayes optimal tilt estimates without distributional assumptions.

Main Results:

• Tilt estimates are moderately accurate, influenced by prior probability distributions.
• Accuracy significantly increases when cues are similar or slant is > 40°.
• Luminance and texture cues can override disparity cues when they agree.

Conclusions:

• Simplifying assumptions in cue combination are often valid for natural scene tilt estimation.
• Moderate accuracy aligns with subjective viewing of small scene patches.
• Improved accuracy in specific conditions supports global surface perception via local estimates.