Seeing Through Satellite Images at Street Views.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Street-view synthesis from satellite imagery is challenging due to sparse views and large viewpoint changes.
• Existing methods struggle with photorealistic rendering and viewpoint consistency.

Purpose of the Study:

• To develop a novel approach for photorealistic street-view panorama synthesis from satellite images.
• To address the challenges of sparse-view learning and extreme viewpoint variations.

Main Methods:

• Learning a satellite image-conditioned neural radiance field from paired satellite and street-view images.
• Modeling street-view specific elements like sky and illumination effects using neural networks.
• Introducing the Sat2Density++ approach for enhanced panorama rendering.

Main Results:

• Sat2Density++ successfully renders photorealistic street-view panoramas.
• The generated panoramas are consistent across multiple views.
• The synthesized images remain faithful to the input satellite imagery.
• Evaluated on urban and suburban datasets, demonstrating robust performance.

Conclusions:

• Sat2Density++ effectively synthesizes high-fidelity street-view panoramas from satellite data.
• The method overcomes key challenges in cross-view synthesis.
• This work advances the state-of-the-art in satellite-to-street-view image generation.