Figure 1: Input image, segmentation mask for real and virtual images. MARVIS outperforms the state-of-the-art by incorporating synthetic training data and robust geometry and motion-aware learning.

MARVIS proposes a novel approach for segmenting real and virtual image regions, essential for marine robotics applications, particularly near the water surface. This paper demonstrates a method combining domain-invariant information, a Motion Entropy Kernel, and Epipolar Geometric Consistency. Our method, trained on synthetic data, achieves over 78% IoU and 86% F1-score on unseen real-world data, providing robust performance with low computational costs. MARVIS offers 43 FPS on a single GPU and ensures efficiency and accuracy in real-virtual image segmentation.

Jiayi Wu, Xiaomin Lin, Shahriar Negahdaripour, Cornelia Fermuller, Yiannis Aloimonos