BiFuse++: Self-Supervised and Efficient Bi-Projection Fusion for 360° Depth Estimation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Monocular 360° depth estimation is crucial for applications like autonomous systems, driven by the increasing use of spherical cameras.
• Existing methods, such as BiFuse, utilize bi-projection fusion but require extensive, costly data collection for training.
• Self-training offers a scalable alternative but lacks integration with advanced fusion techniques.

Purpose of the Study:

• To develop a novel framework, BiFuse++, that integrates bi-projection fusion with self-training for monocular 360° depth estimation.
• To enhance the performance and stability of self-supervised depth estimation from 360° videos.

Main Methods:

• Proposed BiFuse++ framework combining bi-projection fusion with a self-training scheme.
• Introduced a new fusion module to leverage information from diverse projection types.
• Developed a Contrast-Aware Photometric Loss to stabilize self-training on real-world video data.

Main Results:

• Achieved state-of-the-art performance in both supervised and self-supervised experiments on benchmark datasets.
• Demonstrated improved performance and stability of the BiFuse++ framework compared to existing methods.

Conclusions:

• BiFuse++ effectively combines bi-projection fusion and self-training for enhanced monocular 360° depth estimation.
• The proposed methods significantly advance self-supervised learning capabilities in this domain, reducing data collection dependency.