Neural Field Convolutions by Repeated Differentiation
arXiv:2304.01834 · doi:10.1145/3618340
Abstract
Neural fields are evolving towards a general-purpose continuous representation for visual computing. Yet, despite their numerous appealing properties, they are hardly amenable to signal processing. As a remedy, we present a method to perform general continuous convolutions with general continuous signals such as neural fields. Observing that piecewise polynomial kernels reduce to a sparse set of Dirac deltas after repeated differentiation, we leverage convolution identities and train a repeated integral field to efficiently execute large-scale convolutions. We demonstrate our approach on a variety of data modalities and spatially-varying kernels.
References in corpus (9)
- Instant Neural Graphics Primitives with a Multiresolution Hash Encoding
- Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains
- Deep Parametric Continuous Convolutional Neural Networks
- Implicit Neural Representations with Periodic Activation Functions
- Monte Carlo Convolution for Learning on Non-Uniformly Sampled Point Clouds
- Neural Point Catacaustics for Novel-View Synthesis of Reflections
- SAPE: Spatially-Adaptive Progressive Encoding for Neural Optimization
- Spelunking the Deep: Guaranteed Queries on General Neural Implicit Surfaces via Range Analysis
- Polynomial Neural Fields for Subband Decomposition and Manipulation