Structure from stereo by associative learning of the constraints

Area of Science:

• Computational Neuroscience
• Computer Vision
• Machine Learning

Background:

• Stereo vision is crucial for depth perception.
• Existing computational models often require explicit smoothness constraints.
• Associative learning offers a potential mechanism for developing implicit constraints.

Purpose of the Study:

• To propose a computational model of structure from stereo vision.
• To investigate the natural development of smoothness constraints through associative learning.
• To evaluate the model's ability to map disparity data to surface depth.

Main Methods:

• Utilized banks of disparity-selective graded response units as input.
• Employed a pseudo-Markov process to generate surface data (shaded, marked, and random-dot).
• Trained the model using associative learning on example surface mappings.

Main Results:

• The model successfully learned mappings from disparity to depth for trained surfaces.
• It generalized to solve stereo mappings for previously unseen surfaces generated by the same process.
• The model demonstrated the ability to solve some random-dot stereo problems after learning shaded surfaces.

Conclusions:

• Associative learning can naturally induce smoothness constraints in structure from stereo models.
• The proposed model shows promise for robust depth perception across different surface representations.
• This approach offers a biologically plausible mechanism for developing visual processing capabilities.