Tensor4D: Efficient Neural 4D Decomposition for High-Fidelity Dynamic Reconstruction and Rendering

Figure 1. Given 4 sparse static RGB camera views of a dynamic scene (a), our proposed Tensor4D decomposition enables multiview reconstruction to achieve fine-grained geometry reconstruction even on human fingers (b) and temporal-consistent novel view synthesis on a 3D holographic display (c,d,e). The 4 cameras are settled on four conners of the display. The proposed method demonstrates low-cost, portable and highly immersive telepresence experience.