ShadowMover: Automatically Projecting Real Shadows onto Virtual Object.

Area of Science:

• Computer Vision and Graphics
• Artificial Intelligence
• Image Synthesis

Background:

• Integrating 3D virtual objects into real-world images requires realistic shadows for scene coherence.
• Synthesizing consistent shadows, especially those cast by real objects onto virtual ones, is challenging without scene geometry or manual effort.

Purpose of the Study:

• To present the first end-to-end solution for automatically projecting real-world shadows onto virtual objects in outdoor scenes.
• To enable visually realistic composite images for applications like photo editing and augmented reality.

Main Methods:

• Introduced a novel shadow representation: the Shifted Shadow Map, encoding shifted real shadow masks.
• Developed ShadowMover, a Convolutional Neural Network (CNN)-based model, to predict Shifted Shadow Maps and generate shadows.
• Trained the model on a large-scale dataset, requiring no explicit scene geometry or manual intervention.

Main Results:

• The ShadowMover model successfully generates plausible shadows on virtual objects.
• The method is robust across various scene configurations, demonstrating effectiveness without geometric priors.
• Experimental validation confirms the method's capability in creating realistic shadow projections.

Conclusions:

• The proposed ShadowMover system offers a fully automated solution for realistic shadow generation in mixed reality.
• This approach significantly advances the state-of-the-art in image editing and augmented reality by addressing a key visual realism challenge.