Neural Radiance Fields for Dynamic View Synthesis Using Local Temporal Priors

Neural Radiance Fields (NeRF) have demonstrated promising results in synthesizing novel view images from a set of unconstrained captured scenes. One important extension of NeRF is using it on non-rigid reconstruction. Although previous NeRF-based methods for dynamic scene reconstruction have presented visually appealing results, they still often show visual artifacts such as blurry or incorrect geometry of an object. One of the causes is that previous work performs reconstruction directly on the entire video sequence. The global temporal information over the video sequence introduces noise to the network, often leading to a non-optimal canonical space representation of the dynamic scene. In this paper, we present Local Temporal (LT) NeRF, a method to synthesize novel views of dynamic scenes using local temporal priors. Our novel LT module provides the local temporal priors using multi-view stereo sampling, and improves the deformation field reconstruction and hyper-space encoding. Our novel loss functions further supervise the NeRF for better optimization. We evaluate our method with dynamic scenes captured from monocular videos, outperforming the state-of-the-art.