Neural Implicit Surfaces with Frequency-Adaptive Positional Encoding

As the Neural Radiance Fields (NeRF) have achieved significant success in view synthesis, some works attempt to apply volume rendering to 3D reconstruction task. However, these methods use positional encoding with uniform frequency for the whole scenario during training, ignoring different regions of the scene and different training stage. Therefore, we aim to design a frequency-adaptive positional encoding. The adaptivity here refers to two points: first, it is adaptive to scene regions, meaning different regions of each scene have different frequency of positional encoding. Second, it is adaptive to the training process, meaning frequency changes during training to adapt to the convergence of the network. To achieve this adaptivity, we use a neural network to learn the frequency field of scenes, that is, using the network to predict the required frequency for each point during the training process. The introduction of the frequency field allows each scene to obtain an adaptive positional encoding, enabling the neural network to learn the geometric information of scenes in a more flexible and efficient manner. Experimental results show that our approach produces more accurate surfaces compared to baseline methods, especially in scenes with complex geometry.