3D Visual Discomfort Assessment With a Weakly Supervised Graph Convolution Neural Network Based on Inaccurately...

Area of Science:

• Neuroscience
• Computer Science
• Human-Computer Interaction

Background:

• Stereoscopic display technology is limited by visual discomfort, necessitating accurate assessment methods.
• Electroencephalography (EEG) offers objective assessment but is hindered by individual differences causing inaccurately labeled data.
• Current methods often overlook crucial information from the brain's visual cortical areas.

Purpose of the Study:

• To develop a robust method for assessing stereoscopic visual discomfort, overcoming limitations of individual differences and inaccurately labeled data.
• To enhance learning by maximizing the utility of available, albeit imperfect, data.
• To integrate visual cortical information for more effective feature representation.

Main Methods:

• Proposed a weakly supervised graph convolutional neural network for visual discomfort (WSGCN-VD).
• Employed a center correction loss with progressive selection for handling inaccurately labeled data.
• Introduced a feature graph module to capture spatio-temporal representations from visual cortical EEG signals.

Main Results:

• The WSGCN-VD model demonstrated effectiveness across various experimental scenarios.
• The proposed method successfully mitigated the impact of inaccurately labeled data on assessment accuracy.
• The feature graph module effectively integrated visual cortical information for improved representations.

Conclusions:

• The WSGCN-VD offers a promising approach for accurate stereoscopic visual discomfort assessment.
• Weakly supervised learning is effective in handling noisy labels in EEG-based discomfort assessment.
• Integrating visual cortical information enhances the performance of objective visual discomfort evaluation.