Depth Map Upsampling via Multi-Modal Generative Adversarial Network.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Autonomous robots require accurate depth perception for environmental interaction in smart homes and cities.
• Low-resolution depth maps from sensors limit robot performance, especially at depth boundaries.
• Standard super-resolution techniques degrade sharpness and accuracy in depth discontinuity regions.

Purpose of the Study:

• To develop a novel Generative Adversarial Network (GAN)-based framework for depth map super-resolution.
• To improve the preservation of both smooth regions and sharp depth boundaries in super-resolved depth maps.
• To create a model that requires only a depth map for high-resolution output at test time.

Main Methods:

• A Generative Adversarial Network (GAN) framework was designed for depth map super-resolution.
• The model was trained using both color images and depth maps (dual-modality training).
• The framework focuses on preserving fine details and discontinuities during resolution enhancement.

Main Results:

• The proposed GAN-based method successfully preserves smooth areas and sharp edges in depth maps.
• Quantitative and qualitative evaluations demonstrate superior performance compared to existing state-of-the-art models.
• The model achieves high-resolution depth map generation using only the low-resolution depth input during inference.

Conclusions:

• The novel GAN framework offers an effective solution for depth map super-resolution.
• This advancement enhances depth perception capabilities for autonomous robots in intelligent environments.
• The method provides a significant improvement over current depth map enhancement techniques.