Unpaired Underwater Image Synthesis with a Disentangled Representation for Underwater Depth Map Prediction

Area of Science:

• Computer Vision
• Robotics
• Oceanography

Background:

• Underwater depth map estimation is crucial for exploration but faces challenges due to limited paired data and dynamic environments.
• Existing methods struggle to bridge the gap between normal and underwater depth map estimation.
• The lack of paired underwater images hinders the development of effective mapping functions.

Purpose of the Study:

• To develop a novel image-to-image framework for synthesizing diverse underwater images.
• To enable accurate underwater depth map estimation from single images.
• To address the scarcity of paired underwater image data.

Main Methods:

• A novel image-to-image translation framework was developed for underwater image synthesis and depth map estimation.
• Hazy in-air images with depth maps were translated into synthetic underwater images to create a paired dataset.
• A coarse-to-fine network was integrated for precise depth map estimation.
• The framework was further enhanced by translating hazy in-air images into continuously changing underwater images.

Main Results:

• The framework successfully generated a diverse range of synthetic underwater images.
• The method demonstrated high precision in underwater depth map estimation.
• Experimental validation was performed on a real-world underwater RGB-D dataset.

Conclusions:

• The proposed framework effectively overcomes the limitations of paired data scarcity in underwater vision.
• The developed method provides a robust solution for both underwater image synthesis and accurate depth map estimation.
• This approach significantly advances the capabilities for underwater exploration and research.