Progressive Fusion for Unsupervised Binocular Depth Estimation Using Cycled Networks.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Supervised deep learning for monocular depth estimation requires expensive ground truth data.
• Existing methods face limitations due to annotation costs and data requirements.

Purpose of the Study:

• To develop a novel unsupervised deep learning approach for accurate depth map prediction.
• To introduce a new network architecture, the Progressive Fusion Network (PFN), for binocular stereo depth estimation.

Main Methods:

• Proposed a Progressive Fusion Network (PFN) employing a multi-scale refinement strategy for stereo depth estimation.
• Implemented a cycle training approach, acting as data augmentation, by stacking the PFN twice.
• Utilized adversarial learning for joint training of the network architecture.

Main Results:

• The proposed unsupervised method achieves competitive performance against existing deep learning approaches.
• Experiments on KITTI, Cityscapes, and ApolloScape datasets validate the model's effectiveness.
• The cycle approach enhances learning by using both real and synthesized images.

Conclusions:

• The novel unsupervised deep learning approach effectively predicts depth maps without ground truth annotations.
• The Progressive Fusion Network (PFN) and cycle training offer a promising direction for unsupervised stereo depth estimation.
• The method demonstrates strong potential for real-world applications requiring depth perception.