Temporal Interpolation is all You Need for Dynamic Neural Radiance Fields

Figure 1. We propose simple yet effective feature interpolation methods for training neural radiance fields of dynamic scenes based on temporal interpolation. We provide two different feature vector representations, neural representation (top) and grid representation (bottom), both of which are the concatenation of static feature vectors and temporally-interpolated dynamic feature vectors. The neural representation exhibits high-quality rendering performance with small-sized models while the grid representation shows competitive rendering results with astonishingly fast training speed.