Self-Supervised Deep Monocular Depth Estimation With Ambiguity Boosting.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Self-supervised learning (SSL) for depth estimation (DE) from stereo images is crucial for autonomous systems.
• Existing SSL methods often struggle with accuracy and consistency, especially under image transformations.

Purpose of the Study:

• To propose a novel two-stage training strategy with ambiguity boosting for self-supervised single-view depth estimation.
• To improve the accuracy and consistency of generated depth maps using a confidence-guided augmentation loss.

Main Methods:

• A two-stage training strategy involving an auto-encoder for coarse depth prior generation.
• An ambiguity boosting loss function for self-supervision in the second stage, incorporating confidence-guided data augmentation.
• Extension of existing depth estimation networks (t-shaped adaptive kernels, exponential disparity volumes) with the new strategy, creating DBoosterNet-t and DBoosterNet-e.

Main Results:

• DBoosterNets achieve competitive and sometimes superior performance compared to state-of-the-art supervised methods.
• DBoosterNets significantly outperform previous self-supervised monocular DE methods on the KITTI dataset.
• Intensive experiments validate the efficacy of the proposed method for self-supervised monocular DE.

Conclusions:

• The proposed two-stage training strategy with ambiguity boosting effectively enhances self-supervised depth estimation.
• The DBoosterNet models demonstrate strong performance, rivaling supervised approaches and surpassing prior self-supervised methods.
• This work offers a significant advancement in self-supervised monocular depth estimation techniques.