AMENet is a monocular depth estimation network designed for automatic stereoscopic display.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Monocular depth estimation is crucial for autostereoscopic displays but faces challenges in accuracy and robustness within complex environments.
• Existing methods struggle to achieve reliable performance across diverse and intricate visual scenarios.

Purpose of the Study:

• To develop an advanced depth estimation network, AMENet, that integrates Vision Transformer (ViT) and Convolutional Neural Network (CNN) architectures.
• To enhance the accuracy and robustness of monocular depth estimation, particularly for applications in autostereoscopic displays.

Main Methods:

• The proposed AMENet fuses ViT for global semantic feature extraction and CNN for detailed feature processing.
• A weight correction module is incorporated to quantify loss relationships and bolster model robustness.
• Input images are processed as sequences of visual features to leverage ViT's global perception capabilities.

Main Results:

• AMENet demonstrated superior accuracy and robustness compared to existing methods on multiple public datasets.
• The network achieved a notable 4.4% accuracy improvement on the KITTI dataset against the baseline.
• Experimental analysis confirmed the method's effectiveness and stability across various scenarios and complex conditions.

Conclusions:

• AMENet presents a significant advancement in monocular depth estimation, offering high accuracy and robustness.
• The fusion of ViT and CNN architectures provides a powerful approach for tackling complex depth estimation challenges.
• The proposed method is well-suited for demanding applications such as autostereoscopic displays.