AMB3R: Accurate Feed-forward Metric-scale 3D Reconstruction with Backend

TL;DR

AMB3R is a fast, feed-forward 3D reconstruction model that uses a sparse volumetric backend to achieve state-of-the-art accuracy in metric-scale dense 3D reconstruction and can be extended to related tasks without fine-tuning.

Contribution

The paper introduces AMB3R, a novel multi-view dense 3D reconstruction approach that leverages a sparse volumetric backend for improved accuracy and versatility across multiple 3D vision tasks.

Findings

Achieves state-of-the-art performance in camera pose and depth estimation.

Outperforms traditional SLAM and SfM methods on benchmarks.

Can be extended to uncalibrated visual odometry and large-scale SfM without fine-tuning.

Abstract

We present AMB3R, a multi-view feed-forward model for dense 3D reconstruction on a metric-scale that addresses diverse 3D vision tasks. The key idea is to leverage a sparse, yet compact, volumetric scene representation as our backend, enabling geometric reasoning with spatial compactness. Although trained solely for multi-view reconstruction, we demonstrate that AMB3R can be seamlessly extended to uncalibrated visual odometry (online) or large-scale structure from motion without the need for task-specific fine-tuning or test-time optimization. Compared to prior pointmap-based models, our approach achieves state-of-the-art performance in camera pose, depth, and metric-scale estimation, 3D reconstruction, and even surpasses optimization-based SLAM and SfM methods with dense reconstruction priors on common benchmarks.