Residual Vision Transformer and Adaptive Fusion Autoencoders for Monocular Depth Estimation

Area of Science:

• Computer Vision
• Deep Learning
• 3D Sensing

Background:

• Monocular depth estimation is crucial for applications like 3D scene reconstruction and autonomous driving.
• Deep learning advancements have enabled monocular depth estimation to surpass traditional stereo camera systems.
• Accurate depth perception from single-view images remains a significant challenge.

Purpose of the Study:

• To propose an end-to-end supervised monocular depth estimation autoencoder using a single camera.
• To enhance the precision of depth maps through a novel network architecture.
• To improve depth estimation performance for foreground objects.

Main Methods:

• Developed an autoencoder with a mixed convolution neural network and vision transformers encoder.
• Implemented an adaptive fusion decoder for effective feature merging.
• Utilized a human perception-aligned loss function for training.

Main Results:

• The proposed autoencoder effectively predicts depth maps from single-view color images.
• Achieved a 28% increase in the first accuracy rate compared to existing methods.
• Reduced the root mean square error by approximately 27% on the NYU dataset.

Conclusions:

• The developed autoencoder demonstrates high-precision monocular depth estimation capabilities.
• The hybrid encoder and adaptive decoder architecture effectively capture multi-scale features.
• The perception-aligned loss function improves focus on critical foreground object depth.