Abstract
The shape of the contour separating two regions strongly influences judgments of which region is "figure" and which is "ground." Convexity and other figure-ground cues are generally assumed to indicate only which region is nearer, but nothing about how much the regions are separated in depth. To determine the depth information conveyed by convexity, we examined natural scenes and found that depth steps across surfaces with convex silhouettes are likely to be larger than steps across surfaces with concave silhouettes. In a psychophysical experiment, we found that humans exploit this correlation. For a given binocular disparity, observers perceived more depth when the near surface's silhouette was convex rather than concave. Our findings show that convexity should be reclassified as a metric depth cue. They also suggest that the dichotomy between metric and nonmetric depth cues is false and that the depth information provided many cues should be evaluated with respect to natural-scene statistics.
Introduction: The Metric vs. Ordinal Paradox
In classical vision science, depth cues are categorized into two distinct groups. Metric cues, such as binocular disparity or motion parallax, provide a quantitative basis for calculating the precise distance between objects. In contrast, ordinal cues, such as occlusion or convexity, are traditionally thought to only indicate the relative order of objects (e.g., "A is in front of B") without conveying information about the magnitude of the spatial gap.
However, a persistent paradox exists: the shape of an object's silhouette—whether it is convex or concave—significantly alters how much depth a person perceives, even when the primary metric cue (disparity) is held constant. This research investigates whether this phenomenon is a geometric anomaly or a sophisticated adaptation to the statistical regularities of the natural world.
Methodology: Analyzing Natural Scenes and Human Perception
The study employed a three-pronged approach to bridge the gap between environmental statistics and human behavior.
Natural-Scene Analysis and Datasets
The researchers analyzed approximately 450,000 contour points from a dataset combining high-resolution luminance images with co-registered laser-range (LIDAR) data. This allowed for the precise measurement of physical depth steps associated with specific contour shapes in real-world environments. Five participants hand-selected perceptually relevant region-bounding contours to ensure the analysis focused on meaningful boundaries.
Psychophysical Experiments and Modeling
Human observers viewed random-dot stereograms where binocular disparity and contour shape (convex, concave, or straight) were independently manipulated. Participants judged which of two stimuli exhibited a greater perceived depth separation. To interpret the results, the authors developed a Bayesian probabilistic model to test if human depth judgments result from integrating noisy disparity signals with "prior" distributions derived from natural-scene statistics.
Results: Convexity as a Metric Depth Cue
The analysis revealed a fundamental statistical link: in natural scenes, convex silhouettes are significantly more likely to be associated with larger depth steps than concave ones. This correlation likely arises because most discrete objects are convex; a convex contour typically belongs to a foreground object separated from a distant background.
Key findings include:
- Statistical Power Law: Depth distributions in the environment follow a power law ($P \approx \Delta^k$). The difference in the power-law exponent between convex and concave regions was measured at approximately 0.4.
- Perceptual Bias: Convexity significantly biased human depth perception. To make a concave object appear as deep as a convex one, the physical disparity had to be increased by 1.1 to 3.9 arcmin.
- Model Fit: A simplified 5-parameter power-law model provided an excellent fit for the human data, suggesting that the visual system utilizes a streamlined representation of these environmental statistics.
Discussion: Bayesian Integration and Future Directions
These findings challenge the traditional taxonomy of depth cues, suggesting that the distinction between "metric" and "ordinal" is a false dichotomy. Any cue correlating with depth magnitude in the environment can serve as a metric cue.
While the direction of the effect in humans matched natural statistics, the magnitude was notably larger. This discrepancy may be due to the "purity" of experimental stimuli compared to the noisier contours in natural datasets. Future research could explore how the brain integrates convexity with other strong statistical priors, such as the "lower region" cue, and identify the specific neural circuits in the visual cortex responsible for this Bayesian integration.
Summary updated