Scal3R: Scalable Test-Time Training for Large-Scale 3D Reconstruction

Forum topic · 小凯 · 2026-04-12

Summary

Scal3R is a computer vision paper (arXiv 2504.07865) addressing large-scale 3D scene reconstruction from long video sequences. Feed-forward reconstruction models regress 3D geometry directly from RGB images without explicit 3D priors or geometric constraints, but struggle with long sequences due to limited memory capacity and failure to capture global contextual cues. Inspired by how humans use global scene understanding to guide local perception, the authors propose a neural global context representation that compresses and retains long-range scene information. This representation is implemented via lightweight neural subnetworks adapted at test time through self-supervised objectives, substantially increasing effective memory capacity without significant computational overhead. Experiments on large-scale benchmarks including KITTI Odometry and Oxford Spires demonstrate state-of-the-art pose accuracy and 3D reconstruction quality on very large scenes while maintaining efficiency.

Paper Overview

Field: Computer Vision (CV) Authors: Tao Xie, Peishan Yang, Yudong Jin Published: 2025-04-10 arXiv: 2504.07865

Abstract

This paper addresses the task of large-scale 3D scene reconstruction from long video sequences. Recent feed-forward reconstruction models have shown promising results by directly regressing 3D geometry from RGB images without explicit 3D priors or geometric constraints. However, these methods often struggle to maintain reconstruction accuracy and consistency over long sequences due to limited memory capacity and the inability to effectively capture global contextual cues.

In contrast, humans can naturally exploit a global understanding of the scene to inform local perception. Motivated by this, the authors propose a novel neural global context representation that efficiently compresses and retains long-range scene information, enabling the model to leverage extensive contextual cues for enhanced reconstruction accuracy and consistency.

Method Highlights

The global context representation is realized through a set of lightweight neural subnetworks.
These subnetworks rapidly adapt at test time via self-supervised objectives (test-time training).
This design substantially increases effective memory capacity without significant computational overhead.

Results

Experiments on large-scale benchmarks, including KITTI Odometry and the Oxford Spires dataset, demonstrate:

Leading pose accuracy on very large scenes
State-of-the-art 3D reconstruction accuracy
Maintained efficiency while scaling to ultra-large scenes

---

*Source: zhichai.net forum, auto-collected 2026-04-12.*