Image statistics determine the integration of visual cues to motion-in-depth

Area of Science:

• Visual neuroscience
• Perception psychology
• Computational modeling

Background:

• Motion-in-depth perception is vital for survival, aiding hazard avoidance and threat response.
• Humans utilize changing disparity (CD) and changing image size (CS) as key visual cues for depth perception.
• Integrating these cues presents a computational challenge due to reliance on scene parameters like distance, object size, and viewing distance.

Purpose of the Study:

• To investigate how humans integrate visual cues for motion-in-depth perception.
• To determine the role of joint probabilities of scene parameters and cue co-occurrence in cue integration.
• To compare the predictive power of novel probabilistic models against standard linear models for human performance.

Main Methods:

• Developed computational models incorporating sensitivity to joint probabilities of scene parameters and co-occurrence of CD and CS signals.
• Tested model predictions against human performance data in speed-in-depth and cue conflict discrimination tasks.
• Compared performance of probabilistic models with traditional linear integration models.

Main Results:

• Models accounting for joint probabilities and cue co-occurrence accurately predicted human performance in depth perception tasks.
• Standard linear integration models failed to predict human performance under similar conditions.
• Cue integration is influenced by sensory signal uncertainty and the mapping to real-world properties.

Conclusions:

• Human motion-in-depth cue integration is influenced by statistical learning of scene properties and cue relationships.
• Probabilistic cue integration models offer a better explanation of human performance than linear models.
• Scene and image statistics play a crucial role in the perception of motion-in-depth.