LaRa: Efficient Large-Baseline Radiance Fields

TL;DR

LaRa introduces a novel transformer-based approach that combines local and global reasoning for efficient large-baseline radiance field reconstruction, achieving high fidelity and robustness in 3D scene modeling.

Contribution

It unifies local and global reasoning in transformer layers for improved large-baseline radiance field reconstruction, with scene representation as Gaussian Volumes and efficient feed-forward processing.

Findings

High-fidelity 360-degree radiance field reconstruction

Faster convergence compared to previous methods

Robust performance in zero-shot and out-of-domain tests

Abstract

Radiance field methods have achieved photorealistic novel view synthesis and geometry reconstruction. But they are mostly applied in per-scene optimization or small-baseline settings. While several recent works investigate feed-forward reconstruction with large baselines by utilizing transformers, they all operate with a standard global attention mechanism and hence ignore the local nature of 3D reconstruction. We propose a method that unifies local and global reasoning in transformer layers, resulting in improved quality and faster convergence. Our model represents scenes as Gaussian Volumes and combines this with an image encoder and Group Attention Layers for efficient feed-forward reconstruction. Experimental results demonstrate that our model, trained for two days on four GPUs, demonstrates high fidelity in reconstructing 360 deg radiance fields, and robustness to zero-shot and out-of-domain testing. Our project Page: https://apchenstu.github.io/LaRa/.