Semi-Supervised Semantic Segmentation for Light Field Images Using Disparity Information.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Light field (LF) images capture multi-view information, crucial for scene understanding.
• Semantic segmentation of LF images is vital but hindered by the need for extensive pixel-wise annotations in supervised methods.

Purpose of the Study:

• To develop a semi-supervised method for LF semantic segmentation that minimizes the requirement for labeled data.
• To leverage LF disparity information for improved semantic segmentation accuracy.

Main Methods:

• An unsupervised disparity estimation network was designed to generate disparity maps for each view.
• Pseudo-labels were generated for peripheral views using disparity maps and central view labels, with predictions merged for reliability.
• A disparity-semantics consistency loss and a comprehensive contrastive learning scheme (pixel-level and object-level) were introduced.

Main Results:

• The proposed method achieves state-of-the-art performance on benchmark LF semantic segmentation datasets.
• It demonstrates comparable results to fully supervised methods, even with significantly reduced labeled data (e.g., 1/2 protocol).

Conclusions:

• The semi-supervised approach effectively reduces annotation burden in LF semantic segmentation.
• Harnessing LF disparity information and contrastive learning significantly enhances segmentation performance and feature representation.