CLoSeR: Closing the Loop for Long-Context Streaming Reconstruction

Moyang Li; Zihan Zhu; Wei Zhang; Marc Pollefeys; Daniel Barath

cs.CVOct 1, 2026

CLoSeR: Closing the Loop for Long-Context Streaming Reconstruction

Authors: Moyang Li, Zihan Zhu, Wei Zhang, Marc Pollefeys, Daniel Barath

Organizations: ETH Zurich · University of Stuttgart · Microsoft

Abstract

Feedforward foundation models have recently shown remarkable 3D reconstruction capabilities. However, existing models exhibit large tracking drift in long-context streaming reconstruction due to error accumulation. In this paper, we revisit loop closure with streaming reconstruction foundation models to enable accurate, drift-free, kilometer-scale reconstruction. Specifically, our method detects loop candidates through global descriptor retrieval, and constructs loop-conditioned windows to estimate the relative poses between looped frames. Given the observation that our adopted streaming reconstruction backbone produces a globally consistent scale, we optimize all frame poses on the SE(3) manifold with sequential and loop closure constraints, avoiding the pose graph optimization on the Sim(3) or higher-dimensional SL(4) manifolds employed in prior works. Extensive experiments show that our method reduces drift and produces consistent geometry on kilometer-scale sequences, significantly outperforming the state of the art. Code is available at https://github.com/MoyangLi00/CLoSeR.git.

Figures & tables

Figure 2 : System overview. Given a streaming monocular sequence, our method processes it in a sliding-window manner. For each window, we first tokenize the frames with a DINO-based encoder and then apply a stack of attention layers, comprising frame attention, sliding-window attention, and TTT layers, to maintain global consistency across the sequence. Upon the arrival of each new window, we additionally extract SALAD descriptors to identify potential loop closures against prior windows. When such candidates are detected, we construct a loop-conditioned window by pairing $N_{\text{win}}/2$ frames from the current window with $N_{\text{win}}/2$ frames from the corresponding prior window, predict the relative poses between loop edges from a single forward pass, and perform pose graph optimization on the $\mathrm{SE}(3)$ manifold to correct tracking drift.

Panel of Figure 3 — Figure 3 : Reconstruction comparison. LoGeR ( Zhang et al., 2026 ) fails to close loops and produces misaligned duplicate geometry, whereas our method recovers globally consistent reconstructions. Key regions are highlighted with red dashed boxes.

Panel of Table 1 — Table 1 : Comparison of absolute trajectory error (ATE RMSE $\downarrow$ [m]) on VBR ( Brizi et al., 2024 ) . The top and bottom blocks denote optimization-based and feedforward methods, respectively. Optimization-based methods are shown in gray , with the best result per sequence in bold . Among feedforward methods, best results are highlighted as first , second , and third .

Panel of Table 2 — Table 2 : Comparison of absolute trajectory error (ATE RMSE $\downarrow$ [m]) on KITTI ( Geiger et al., 2013 ) . Traditional deep visual SLAM systems suffer from scale drift under KITTI’s forward-dominant motion. Our approach outperforms all baselines.

Panel of Table 3 — Table 3 : Comparison of absolute trajectory error (ATE RMSE $\downarrow$ [m]) on DROID-W ( Li et al., 2026 ) . Optimization-based methods are shown in gray , with the best result per sequence in bold . Among feedforward methods, best results are highlighted as first , second , and third . Our approach yields robust and accurate camera pose estimation on dynamic sequences, despite not explicitly handling dynamic objects.

Panel of Table 4 — Table 4 : Comparison of absolute trajectory error (ATE RMSE $\downarrow$ [m]) on Oxford Spires ( Tao et al., 2024 ) . Our approach achieves the best pose estimation performance on this challenging dataset, demonstrating the superiority and robustness of our method.

Panel of Figure 4 — Figure 4 : Trajectory comparison on KITTI Odometry ( Geiger et al., 2013 ) , VBR ( Brizi et al., 2024 ) , DROID-W ( Li et al., 2026 ) , and Oxford Spires ( Tao et al., 2024 ) datasets. Groundtruth trajectories ( coral , dashed) are overlaid with the estimates from each method ( gray-blue , solid) after Sim(3) Umeyama alignment ( Umeyama, 1991 ) .

Method	ATE RMSE [m] $\downarrow$
(a) LoGeR (w/o Loop Closure)	36.16
(b) Neighbor-only constr.	14.25
(c) w/o Huber loss	14.15
(d) Sim(3) optimization	13.33
(e) SL(4) optimization	20.36
Full method	11.15

Table 5 : Ablation study. Average tracking error on the VBR ( Brizi et al., 2024 ) dataset. We ablate five components of our method and compare against the full model.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Table 6 — Table 6 : Runtime and GPU memory analysis. We report the average runtime breakdown (as a percentage of total runtime), the average number of windows, FPS, peak GPU memory usage, and trajectory accuracy on the KITTI ( Geiger et al., 2013 ) dataset. PGO denotes pose graph optimization.

Method	Relative Scale Error Regular Window $\downarrow$	Relative Scale Error LC Window $\downarrow$
VGGT-long ( Deng et al., 2025 )	18.8%	49.3%
VGGT-SLAM 2.0 ( Maggio and Carlone, 2026 )	32.4%	147.2%
CLoSeR (Ours)	5.8%	6.5%

Appendix

Table 7 : Analysis of cross-window scale consistency. We report average relative scale error on the KITTI ( Geiger et al., 2013 ) dataset.

Panel of Figure 5 — Figure 5 : Point clouds comparison. Our method delivers more accurate geometry, while LoGeR exhibits noticeable ghosting artifacts caused by inaccurate pose estimation.

Panel of Figure 6 — Figure 6 : Additional trajectory comparison. Groundtruth trajectories ( coral , dashed) are overlaid with the estimates from each method ( gray-blue , solid) after Sim(3) Umeyama alignment ( Umeyama, 1991 ) . Our approach robustly yields the most accurate trajectories across all sequences.

Explore similar work

LingBot-Map: Geometric Context Transformer for Streaming 3D Reconstruction
Apr 15, 2026Lin-Zhuo Chen, Jian Gao, Shangzhan Zhang +9Feed-Forward 3D Reconstruction 3D Reconstruction
Glob3R: Global Structure-from-Motion with 3D Foundation Models
Jul 10, 2026Junyuan Deng, Heng Li, Kejie Qiu +7Feed-Forward 3D Reconstruction Structure-From-Motion
The Hong Kong University of Science and Technology · Tongyi Lab, Alibaba Group · Nanjing University +1
HorizonStream: Long-Horizon Attention for Streaming 3D Reconstruction
May 22, 2026Chong Cheng, Peilin Tao, Nanjie Yao +93D Reconstruction Shape Correspondence
HKUST(GZ) · Horizon Robotics · CASIA +2

Apr 15, 2026, cs.CVJ/K move · Enter open · S save

LingBot-Map: Geometric Context Transformer for Streaming 3D Reconstruction

Lin-Zhuo Chen, Jian Gao, Shangzhan Zhang, Yihang Chen+8

Feed-Forward 3D Reconstruction 3D Reconstruction Simultaneous Localization And Mapping Trajectory Generation

Open paper AskPDF arXiv

Streaming 3D reconstruction aims to recover 3D information, such as camera poses and point clouds, from a video stream, which necessitates geometric accuracy, temporal consistency, and computational efficiency. Motivated by the principles of Simultaneous Localization and Mapping (SLAM), we introduce LingBot-Map, a feed-forward 3D foundation model for reconstructing scenes from streaming data, built upon a geometric context transformer (GCT) architecture. A defining aspect of LingBot-Map lies in its carefully designed attention mechanism, which integrates an anchor context, a pose-reference window, and a trajectory memory to address coordinate grounding, dense geometric cues, and long-range drift correction, respectively. This design keeps the streaming state compact while retaining rich geometric context, enabling stable efficient inference at around 20 FPS on 518 x 378 resolution inputs over long sequences exceeding 10,000 frames. Extensive evaluations across a variety of benchmarks demonstrate that our approach achieves superior performance compared to both existing streaming and iterative optimization-based approaches.

Abstract

Explore similar work

LingBot-Map: Geometric Context Transformer for Streaming 3D Reconstruction

Glob3R: Global Structure-from-Motion with 3D Foundation Models

HorizonStream: Long-Horizon Attention for Streaming 3D Reconstruction