PointForward: Feedforward Driving Reconstruction through Point-Aligned Representations

TL;DR

PointForward introduces a novel feedforward 3D scene reconstruction method for autonomous driving that explicitly models cross-view consistency and dynamic instances using point-aligned representations and scene graphs.

Contribution

It proposes a new point-aligned, feedforward reconstruction framework with explicit cross-view and instance-level dynamic modeling, improving over pixel-aligned methods.

Findings

Achieves state-of-the-art results on large-scale driving benchmarks.

Enforces explicit cross-view consistency in a single feedforward pass.

Effectively models dynamic scene elements with scene graphs and 3D bounding boxes.

Abstract

High-fidelity reconstruction of driving scenes is crucial for autonomous driving. While recent feedforward 3D Gaussian Splatting (3DGS) methods enable fast reconstruction, their per-pixel Gaussian prediction paradigm often suffers from multi-view inconsistency and layering artifacts. Moreover, existing methods often model dynamic instances via dense flow prediction, which lacks explicit cross-view correspondence and instance-level consistency. In this paper, we propose PointForward, a feedforward driving reconstruction framework through point-aligned representations. Unlike pixel-aligned methods, we initialize sparse 3D queries in world space and aggregate multi-view image information via spatial-temporal fusion onto these queries, enforcing explicit cross-view consistency in a single feedforward pass. To handle scene dynamics, we introduce scene graphs that explicitly organize moving instances during reconstruction. By leveraging 3D bounding boxes, our method enables instance-level motion propagation and temporally consistent dynamic representations. Extensive experiments demonstrate that PointForward achieves state-of-the-art performance on large-scale driving benchmarks. The code will be available upon the publication of the paper.