Just Noticeable Differences in 3D Shape Perception: A Measure of Estimation Noise or Cue Gain?

Abstract

The present study sought to test two fundamentally distinct interpretations of the Just Noticeable Difference (JND) for depth discrimination. According to the widely accepted Bayesian Maximum Likelihood Estimation (MLE) model of 3D processing, the JND measures the standard deviation of the noise corrupting a depth estimate. The Intrinsic Constraint (IC) model of cue integration alternatively suggests that the JND is the result of task related decision noise rather than noise directly corrupting the depth estimate. According to this account, the JND is inversely proportional to the slope, or gain, of the perceptual function relating physical depth to perceived depth. To test this novel interpretation, we tested depth discrimination with a classic 2-Interval Forced Choice task in which the gains of a fixed standard stimulus and varying comparison stimulus could either have a High or Low value, or, according to Bayesian models, High or Low reliabilities. Bayesian models predict that the JND should depend on both the reliability of the standard and comparison stimuli. However, according to the IC model, the JND only depends on the gain of the comparison stimulus, and therefore there should be no difference between conditions where the comparison has the same gain. Empirical results with texture stimuli closely align with the IC predictions. In conclusion, since JNDs do not measure the magnitude of the noise of 3D estimates as assumed by the Bayesian approach, these findings lead to a radically different interpretation of previous depth discrimination data.