Stereo matching via self-supervised pre-training with view synthesis-based pretext task.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Deep learning has advanced stereo matching.
• Current methods require extensive labeled data, limiting practical use.
• A need exists for stereo matching techniques less dependent on labeled datasets.

Purpose of the Study:

• To propose a self-supervised pre-training method (SPVP) for stereo matching.
• To reduce the dependency on large-scale labeled images for training.
• To improve the applicability of deep learning-based stereo matching.

Main Methods:

• A novel view synthesis-based pretext task is introduced.
• A monocular-part network converts right views to pseudo-left views.
• A stereo-part network learns correspondences between image pairs using pre-training.

Main Results:

• The proposed SPVP method effectively learns relationships between image pairs.
• Pre-training with SPVP enables stereo matching networks to train with limited labeled data.
• Experimental results validate the effectiveness of the SPVP approach.

Conclusions:

• SPVP significantly alleviates the need for large labeled datasets in stereo matching.
• The view synthesis pretext task successfully transfers learned relationships to the main stereo matching task.
• The proposed method enhances the efficiency and applicability of deep learning for stereo matching.