Scale Propagation Network for Generalizable Depth Completion.

Area of Science:

• Computer Vision
• Deep Learning
• 3D Perception

Background:

• Depth completion is vital for 3D perception but current deep learning models struggle to generalize to new scenes.
• Conventional normalization layers in these models hinder scale estimation, a key factor for generalization.

Purpose of the Study:

• To develop a novel normalization method that improves the generalization of depth completion models.
• To introduce a new network architecture that leverages this normalization for superior performance.

Main Methods:

• Proposed a scale propagation normalization (SP-Norm) method that propagates scale information from input to output.
• Developed a new network architecture using SP-Norm with the ConvNeXt V2 backbone.
• Trained and evaluated the model on diverse datasets with various sparse depth map types.

Main Results:

• The proposed SP-Norm method effectively preserves scale information, unlike conventional normalization layers.
• The new network architecture achieved superior accuracy in depth completion across six unseen datasets.
• The model demonstrated faster inference speed and lower memory usage compared to state-of-the-art methods.

Conclusions:

• The SP-Norm method is a key innovation for overcoming the generalization limitations in deep learning-based depth completion.
• The developed architecture offers a promising direction for robust and efficient 3D perception systems.
• This work advances the capability of AI models to accurately interpret 3D environments from sparse data.