Abstract
Recent methods for neural surface representation and rendering, for exampleNeuS, have demonstrated remarkably high-quality reconstruction of staticscenes. However, the training of NeuS takes an extremely long time (8 hours),which makes it almost impossible to apply them to dynamic scenes with thousandsof frames. We propose a fast neural surface reconstruction approach, calledNeuS2, which achieves two orders of magnitude improvement in terms ofacceleration without compromising reconstruction quality. To accelerate thetraining process, we integrate multi-resolution hash encodings into a neuralsurface representation and implement our whole algorithm in CUDA. We alsopresent a lightweight calculation of second-order derivatives tailored to ournetworks (i.e., ReLU-based MLPs), which achieves a factor two speed up. Tofurther stabilize training, a progressive learning strategy is proposed tooptimize multi-resolution hash encodings from coarse to fine. In addition, weextend our method for reconstructing dynamic scenes with an incrementaltraining strategy. Our experiments on various datasets demonstrate that NeuS2significantly outperforms the state-of-the-arts in both surface reconstructionaccuracy and training speed. The video is available athttps://vcai.mpi-inf.mpg.de/projects/NeuS2/ .