Adaptive Context-Aware Multi-Modal Network for Depth Completion.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Depth completion recovers dense depth maps from sparse data and RGB images.
• Standard convolutions struggle with sparse depth data, limiting context modeling.
• Existing methods lack effective multi-modal feature fusion for depth completion.

Purpose of the Study:

• To develop an effective depth completion method using graph propagation and attention mechanisms.
• To improve the modeling of observed spatial contexts from sparse depth data.
• To enhance multi-modal feature fusion for accurate depth map recovery.

Main Methods:

• Constructing multi-scale graphs from observed pixels for spatial context.
• Applying attention mechanisms to graph propagation for adaptive context modeling.
• Utilizing graph propagation on RGB and depth modalities separately, followed by symmetric gated fusion.

Main Results:

• The proposed Adaptive Context-Aware Multi-Modal Network (ACMNet) achieves state-of-the-art performance on KITTI and NYU-v2 benchmarks.
• ACMNet demonstrates superior depth completion accuracy compared to existing methods.
• The model achieves these results with a reduced number of parameters.

Conclusions:

• Graph propagation with attention is effective for modeling spatial contexts in depth completion.
• Symmetric gated fusion efficiently integrates multi-modal features.
• ACMNet offers a parameter-efficient and high-performance solution for depth completion.