NoDrift3R

NoDrift3R: Raymap-Guided Coupling for Drift-Robust Unposed Feed-Forward 3D Reconstruction

Project Page | Paper | Code

<p align="center"> <img src="https://xiangyu1sun.github.io/NoDrift3R-project-page/static/images/problem.png" alt="NoDrift3R overview" width="100%"> </p>

NoDrift3R is an ECCV 2026 method for drift-robust, pose-free feed-forward 3D Gaussian Splatting reconstruction. It targets cumulative camera-pose drift in long, unposed image sequences by explicitly coupling geometry and appearance. Gaussian centers are anchored to raymap-induced geometry, while RGB reconstruction, raymap consistency, and camera regularization are optimized jointly to improve rendering, geometry, and pose estimation together.

Highlights

• Drift-robust long-sequence reconstruction: directly addresses cumulative camera-pose errors that degrade geometry and rendering quality.
• Raymap-Guided Coupling: connects Gaussian geometry with raymap predictions to reduce pose-geometry ambiguity.
• Geometry-appearance synergy: jointly optimizes RGB reconstruction, raymap consistency, and camera regularization in a unified objective.
• Dual-Frequency Viewpoint Scheduling: combines easy-to-hard interval expansion with replay of short-interval image pairs for stable training across temporal ranges.
• Pose-free feed-forward inference: reconstructs scenes and estimates camera poses without requiring input camera poses.

Checkpoints

This repository contains the following checkpoints:

• dl3dv_giant_2_8_H20.ckpt
• dl3dv_large_4_8_5090.ckpt
• mix_dataset_giant_4_8_H20.ckpt
• mix_dataset_large_8_8_5090.ckpt

Please follow the GitHub repository and project page for code, configuration, dataset, and usage updates.

Citation

If you find NoDrift3R useful in your research, please cite:

@misc{sun2026nodrift3rraymapguidedcouplingdriftrobust,
  title         = {NoDrift3R: Raymap-Guided Coupling for Drift-Robust Unposed Feed-Forward 3D Reconstruction},
  author        = {Xiangyu Sun and Liu Liu and Seungkwon Yang and Jingbing Han and Seungtae Nam and Zhizhong Su and Eunbyung Park},
  year          = {2026},
  eprint        = {2607.07168},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2607.07168}
}