Likelihood-Based Generative Radiance Field with Latent Space Energy-Based Model for 3D-Aware Disentangled Image Representation

TL;DR

This paper introduces NeRF-LEBM, a likelihood-based 3D-aware generative model that combines Neural Radiance Fields with energy-based priors to produce and understand 2D images from 3D representations, even with incomplete data.

Contribution

The paper presents a novel likelihood-based framework integrating NeRF with energy-based models for 3D-aware image generation and inference, accommodating known and unknown camera poses.

Findings

NeRF-LEBM can infer 3D structures from 2D images.

It can generate images with novel views and objects.

It learns effectively from incomplete or pose-unknown images.

Abstract

We propose the NeRF-LEBM, a likelihood-based top-down 3D-aware 2D image generative model that incorporates 3D representation via Neural Radiance Fields (NeRF) and 2D imaging process via differentiable volume rendering. The model represents an image as a rendering process from 3D object to 2D image and is conditioned on some latent variables that account for object characteristics and are assumed to follow informative trainable energy-based prior models. We propose two likelihood-based learning frameworks to train the NeRF-LEBM: (i) maximum likelihood estimation with Markov chain Monte Carlo-based inference and (ii) variational inference with the reparameterization trick. We study our models in the scenarios with both known and unknown camera poses. Experiments on several benchmark datasets demonstrate that the NeRF-LEBM can infer 3D object structures from 2D images, generate 2D images with novel views and objects, learn from incomplete 2D images, and learn from 2D images with known or unknown camera poses.