Add to Favourites
To login click here
Implicit neural representations (INRs) or neural fields are coordinate-based neural networks that map 3D coordinates to color and density values in 3D space. Recently, neural fields have gained traction in computer vision for representing signals like pictures, 3D shapes/scenes, movies, music, medical images, and weather data. A new study by DeepMind and the University of Haifa presents a strategy for expanding the applicability of functa, a deep learning framework for neural fields, to more extensive and intricate data sets. As an extension of functa, spatial functa replaces flat latent vectors with spatially ordered representations of latent variables, allowing for more sophisticated architectures to be used for downstream tasks.
