Disparity estimation by pooling evidence from energy neurons

Area of Science:

• Computational neuroscience
• Computer vision
• Machine learning

Background:

• Disparity estimation is crucial for 3D perception.
• Existing biologically plausible models often lack probabilistic interpretation.
• The role of disparity energy neuron responses in evidence accumulation is debated.

Purpose of the Study:

• To propose a biologically plausible algorithm for disparity estimation.
• To develop a model with probabilistic output interpretation.
• To improve upon existing disparity estimation techniques.

Main Methods:

• Utilized Bayes factor from statistical hypothesis testing to analyze neuron responses.
• Formulated a probabilistic interpretation of normalized disparity neuron responses.
• Developed a disparity estimation algorithm using biologically plausible operations and pooled normalized responses.

Main Results:

• Demonstrated that normalized disparity neuron responses, not raw responses, represent evidence.
• Showcased that information from different orientation channels can be combined.
• The proposed algorithm outperformed a previous coarse-to-fine model on real stereograms.

Conclusions:

• The developed algorithm offers a probabilistically interpretable and biologically plausible approach to disparity estimation.
• The model's probabilistic nature allows for the identification of occluded or erroneous disparity pixels.
• This work advances the understanding of neural mechanisms underlying stereoscopic vision.