Self-supervised recurrent depth estimation with attention mechanisms.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Depth estimation is crucial for autonomous driving.
• Self-supervised methods estimate depth using camera motion, avoiding ground truth data.
• Current self-supervised methods primarily focus on single-frame estimation.

Purpose of the Study:

• To improve self-supervised monocular depth estimation by leveraging temporal information.
• To investigate the integration of recurrent blocks and attention mechanisms into existing pipelines.
• To propose novel neural network architectures for enhanced depth prediction.

Main Methods:

• Integrating recurrent neural networks (RNNs) and attention mechanisms into a self-supervised depth estimation framework.
• Utilizing temporal information from sequential frames to refine depth predictions.
• Developing new neural network architectures tailored for self-supervised monocular depth estimation.

Main Results:

• The proposed modifications effectively exploit temporal information for depth prediction.
• Experiments on the KITTI dataset demonstrate significant improvements in depth estimation quality.
• The integration of recurrent and attention mechanisms enhances the precision of self-supervised depth models.

Conclusions:

• Temporal information significantly boosts the performance of self-supervised depth estimation models.
• Recurrent and attention mechanisms are effective tools for integrating sequential data in depth prediction.
• The developed architectures offer a promising direction for advancing monocular depth estimation in computer vision.