Joint Soft-Hard Attention for Self-Supervised Monocular Depth Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Self-supervised monocular depth estimation offers a cost-effective alternative to laser sensors for acquiring depth information.
• Current methods, while providing dense depth maps, require improved accuracy for critical applications such as autonomous driving and robot perception.

Purpose of the Study:

• To enhance the accuracy of self-supervised monocular depth estimation.
• To introduce innovative attention mechanisms for improved feature extraction and depth prediction fusion.

Main Methods:

• Integration of a soft attention module to bolster spatial and channel-wise feature extraction with minimal parameter increase.
• Application of a hard attention strategy for superior fusion of multi-scale depth predictions, diverging from traditional fusion techniques.

Main Results:

• The proposed method demonstrates superior performance in self-supervised monocular depth estimation.
• Achieved state-of-the-art results on the standard KITTI benchmark and the Make3D dataset.

Conclusions:

• The combination of soft and hard attention significantly advances self-supervised monocular depth estimation.
• The developed approach offers a promising solution for accurate depth perception in real-world applications.