GRF: Learning a General Radiance Field for 3D Representation and Rendering

TL;DR

This paper introduces a neural network that implicitly models 3D objects and scenes as a general radiance field, enabling high-quality rendering from 2D images with learned local features and attention mechanisms.

Contribution

The method uniquely combines local pixel feature learning with projection to 3D and an attention mechanism for multi-view aggregation, advancing 3D scene representation and rendering.

Findings

Produces realistic novel view synthesis for unseen objects and categories

Handles complex real-world scenes effectively

Outperforms previous methods in quality and generality

Abstract

We present a simple yet powerful neural network that implicitly represents and renders 3D objects and scenes only from 2D observations. The network models 3D geometries as a general radiance field, which takes a set of 2D images with camera poses and intrinsics as input, constructs an internal representation for each point of the 3D space, and then renders the corresponding appearance and geometry of that point viewed from an arbitrary position. The key to our approach is to learn local features for each pixel in 2D images and to then project these features to 3D points, thus yielding general and rich point representations. We additionally integrate an attention mechanism to aggregate pixel features from multiple 2D views, such that visual occlusions are implicitly taken into account. Extensive experiments demonstrate that our method can generate high-quality and realistic novel views for novel objects, unseen categories and challenging real-world scenes.