Intrinsic Omnidirectional Video Decomposition

Rong-Kai Xu, Fang-Lue Zhang, Lei Zhang

Abstract: Intrinsic decomposition of omnidirectional video is a challenging task. We propose a method that can provide temporally consistent decomposition results. Leveraging the 360-degree scene representation, we maintain the global point cloud to propagate and reuse the similar inter-frame content and establish temporal constraints which elevate the quality of frame-wise decomposition while maintaining inter-frame coherence. By optimizing the proposed objective function, our method achieves a precise separation of reflectance and shading components. Comprehensive experiments demonstrate that our approach outperforms existing intrinsic decomposition methods. Our method also hold promise for various video manipulation applications.