Info3R: Information-Adaptive Test-Time Training for 3D Reconstruction

Sunghyun Baek; Hanna Bae; Minchan Kwon; Junmo Kim

cs.CVSep 18, 2026

Info3R: Information-Adaptive Test-Time Training for 3D Reconstruction

Authors: Sunghyun Baek, Hanna Bae, Minchan Kwon, Junmo Kim

Organizations: Department of Electrical Engineering, Korea Advanced Institute of Science and Technology, South Korea

Abstract

Transformer-based models have recently achieved strong performance on 3D reconstruction from images, and recent works extend them to process video streams in an online manner for real-world deployment. However, existing methods overlook two key signals when handling long image streams: the importance of each incoming frame and the information saturation of the model's internal state. In this paper, we propose Info3R, a novel information-adaptive test-time training method for the online 3D reconstruction. We introduce an information-aware state update that modulates the state update strength based on the redundancy and informativeness of each incoming frame. To restore the state's plasticity -- its capacity to incorporate new observations -- we propose a dynamic state reset, triggered by the cumulative magnitude of state updates and the model's prediction confidence and accompanied by an anchor-to-world alignment. Our method achieves consistent improvements on camera pose estimation, video depth estimation, and 3D reconstruction, while substantially mitigating the performance degradation in the long sequence evaluation. Notably, on KITTI Odometry, our method achieves on average 1.68x lower ATE than LongStream, demonstrating its robustness on extended outdoor sequences.

Figures & tables

Fig. 1 : 3D reconstruction on KITTI Odometry Seq. 07. TTT3R fails to estimate the overall trajectory, and LongStream improves over it but fails at loop closure when revisiting the same location. In contrast, Info3R reconstructs a coherent and geometrically consistent scene, demonstrating robustness to long-horizon streaming inputs.

Fig. 2 : Overview of Info3R. For each incoming frame, the Information-Aware State Update modulates the state update with an adaptive learning rate $\alpha_{t}$ derived from information redundancy and informativeness. The Dynamic State Reset re-initializes the saturated state when the trigger metric $\Gamma_{t}$ exceeds a threshold, and preserves global geometry through anchor-to-world alignment.

Methods	KITTI Odometry [ 17 ] ( ATE $\downarrow$ )											Avg. ( $\downarrow$ )
Methods	00 4542x, 3.7km	01 1101x, 2.5km	02 4661x, 5.1km	03 801x, 0.6km	04 271x, 0.4km	05 2761x, 2.2km	06 1101x, 1.2km	07 1101x, 0.7km	08 4071x, 3.2km	09 1591x, 1.7km	10 1201x, 0.9km	Avg. ( $\downarrow$ )
CUT3R	187.79	638.10	279.32	153.64	22.98	152.89	132.52	73.47	233.12	176.99	185.36	203.29
TTT3R	161.60	537.34	263.74	106.43	11.91	145.46	128.47	69.33	236.65	181.62	127.32	179.08
LongStream	77.05	49.45	169.20	3.66	2.18	75.22	12.83	15.66	61.60	77.38	19.34	51.24
Info3R (Ours)	30.03	97.20	86.57	9.11	6.26	23.60	11.21	7.64	29.99	22.89	10.76	30.48

TABLE I : Quantitative comparison on the KITTI Odometry dataset.

Fig. 3 : Camera trajectories on Seq. 00 in KITTI Odometry.

Method	TUM									ScanNet
	Short (50)			Medium (500)			Long (1000)			Short (50)			Medium (500)			Long (1000)
	ATE( $\downarrow$ )	RPE ${}_{\text{T}}$ ( $\downarrow$ )	RPE ${}_{\text{R}}$ ( $\downarrow$ )	ATE( $\downarrow$ )	RPE ${}_{\text{T}}$ ( $\downarrow$ )	RPE ${}_{\text{R}}$ ( $\downarrow$ )	ATE( $\downarrow$ )	RPE ${}_{\text{T}}$ ( $\downarrow$ )	RPE ${}_{\text{R}}$ ( $\downarrow$ )	ATE( $\downarrow$ )	RPE ${}_{\text{T}}$ ( $\downarrow$ )	RPE ${}_{\text{R}}$ ( $\downarrow$ )	ATE( $\downarrow$ )	RPE ${}_{\text{T}}$ ( $\downarrow$ )	RPE ${}_{\text{R}}$ ( $\downarrow$ )	ATE( $\downarrow$ )	RPE ${}_{\text{T}}$ ( $\downarrow$ )	RPE ${}_{\text{R}}$ ( $\downarrow$ )
CUT3R	0.026	0.009	0.327	0.151	0.008	0.417	0.183	0.008	0.544	0.045	0.018	0.478	0.674	0.034	1.161	0.825	0.035	1.203
TTT3R	0.015	0.008	0.310	0.073	0.011	0.373	0.117	0.010	0.480	0.033	0.017	0.450	0.271	0.036	1.096	0.394	0.039	1.162
Info3R (Ours)	0.014	0.008	0.305	0.036	0.010	0.377	0.061	0.011	0.508	0.033	0.018	0.489	0.141	0.022	0.596	0.176	0.023	0.595

TABLE II : Camera Pose estimation across different sequence lengths. For each setting,

\text{RPE}_{\text{T}}

(

\text{RPE}_{\text{trans}}

) and

\text{RPE}_{\text{R}}

(

\text{RPE}_{\text{rot}}

) are reported in the second and third columns of each block, respectively.

Method	Bonn						KITTI
	Short (50)		Medium (250)		Long (500)		Short (50)		Medium (250)		Long (500)
	Abs Rel $(\downarrow)$	$\delta_{1.25}(\uparrow)$	Abs Rel $(\downarrow)$	$\delta_{1.25}(\uparrow)$	Abs Rel $(\downarrow)$	$\delta_{1.25}(\uparrow)$	Abs Rel $(\downarrow)$	$\delta_{1.25}(\uparrow)$	Abs Rel $(\downarrow)$	$\delta_{1.25}(\uparrow)$	Abs Rel $(\downarrow)$	$\delta_{1.25}(\uparrow)$
CUT3R	0.106	0.897	0.104	0.882	0.099	0.907	0.112	0.876	0.129	0.851	0.151	0.806
TTT3R	0.087	0.946	0.103	0.900	0.100	0.922	0.105	0.893	0.112	0.896	0.132	0.866
Info3R (Ours)	0.071	0.966	0.073	0.960	0.077	0.955	0.115	0.869	0.107	0.891	0.116	0.881

TABLE III : Video Depth estimation across different sequence lengths.

Fig. 4 : 3D Reconstruction Quality on 7-Scenes.

Method	NRGBD									7-Scenes
	Short (50)			Medium (200)			Long (400)			Short (50)			Medium (200)			Long (400)
	Acc $(\downarrow)$	Comp $(\downarrow)$	NC $(\uparrow)$	Acc $(\downarrow)$	Comp $(\downarrow)$	NC $(\uparrow)$	Acc $(\downarrow)$	Comp $(\downarrow)$	NC $(\uparrow)$	Acc $(\downarrow)$	Comp $(\downarrow)$	NC $(\uparrow)$	Acc $(\downarrow)$	Comp $(\downarrow)$	NC $(\uparrow)$	Acc $(\downarrow)$	Comp $(\downarrow)$	NC $(\uparrow)$
CUT3R	0.041	0.016	0.673	0.133	0.032	0.600	0.314	0.109	0.554	0.021	0.019	0.607	0.087	0.049	0.564	0.163	0.102	0.533
TTT3R	0.034	0.015	0.680	0.063	0.013	0.625	0.141	0.071	0.591	0.018	0.019	0.610	0.027	0.023	0.581	0.049	0.026	0.558
Info3R (Ours)	0.031	0.014	0.665	0.043	0.009	0.619	0.072	0.018	0.614	0.018	0.018	0.609	0.019	0.021	0.579	0.020	0.020	0.559

TABLE IV : 3D reconstruction across different sequence lengths.

Panel of Figure 9 — Fig. 5 : Analyses on the TUM dynamics dataset.

Method	Bonn (Length 500)		KITTI (Length 500)
Method	Abs Rel( $\downarrow$ )	$\delta<{1.25}(\uparrow)$	Abs Rel( $\downarrow$ )	$\delta<{1.25}(\uparrow)$
Motion Blur (level 5)
CUT3R	0.097	91.8	0.159	77.8
TTT3R	0.091	94.5	0.141	83.2
Info3R w/o $w_{\text{info}}$	0.086	93.8	0.133	88.3
Info3R (Ours)	0.081	94.9	0.132	83.5

TABLE V : Robustness evaluation under image-level corruptions on Bonn and KITTI.

Fig. 6 : Camera trajectories and reset timings.

Method	Mem. (GB, $\downarrow$ )	FPS ( $\uparrow$ )	ATE ( $\downarrow$ )
CUT3R	3.23	15.27	188.49
TTT3R	4.70	15.17	180.00
LongStream	10.77	20.47	77.05
Info3R (Ours)	3.24	13.27	30.03

TABLE VI : Efficiency comparison on KITTI Odometry sequence “00”.

Variants	ATE ( $\downarrow$ )	$\text{RPE}_{\text{rot}}$ ( $\downarrow$ )	$\text{RPE}_{\text{trans}}$ ( $\downarrow$ )
w/o $w_{\text{redundant}}$	45.03	0.239	0.213
w/o $w_{\text{info}}$	32.47	0.283	0.430
Reset w/o $\alpha_{t}$	36.68	0.347	0.490
Reset w/o $\bar{C}_{t}^{\text{world}}$	36.18	0.393	0.932
Info3R (Ours)	30.48	0.333	0.663

TABLE VII : Ablation study on KITTI Odometry.

Task	Dataset	$\gamma$	$\alpha_{\text{base}}$
Long sequence pose estimation	KITTI Odometry	3.2	1.7
Pose estimation	ScanNet	3	1.0
Pose estimation	TUM	20	1.0
Depth estimation	Bonn	3	1.0
Depth estimation	KITTI	0.3	0.4
3D reconstruction	7-Scenes	25	0.4

TABLE VIII : Hyperparameters of our method on each dataset.

\gamma

: reset trigger threshold;

\alpha_{\text{base}}

: base learning rate.

\tau=100

across all datasets.

Explore similar work

BAT3R: Robust Online 3D Reconstruction with Bayesian Adaptive State Updates
Oct 12, 2026Lijian Li, Yuanpeng He, Yudian Zheng +6Streaming 3D Reconstruction Bayesian Filtering
University of Macau, Macau SAR, China · School of Computer Science, Peking University, Beijing, China · Key Laboratory of High Confidence Software Technologies (Peking University), Ministry of Education, Beijing, China
TTSA3R: Training-Free Temporal-Spatial Adaptive Persistent State for Streaming 3D Reconstruction
Jan 30, 2026Zhijie Zheng, Xinhao Xiang, Jiawei ZhangStreaming 3D Reconstruction 3D Reconstruction
University of California, Davis · CA 95616, USA
PAS3R: Pose-Adaptive Streaming 3D Reconstruction for Long Video Sequences
Mar 22, 2026Lanbo Xu, Liang Guo, Caigui Jiang +1Streaming 3D Reconstruction Monocular 3D Reconstruction
State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, China · University of East Anglia, Norwich, NR47TJ, United Kingdom

Oct 12, 2026J/K move · Enter open · S save

BAT3R: Robust Online 3D Reconstruction with Bayesian Adaptive State Updates

Lijian Li, Yuanpeng He, Yudian Zheng, Chung-ju Huang+5

University of Macau, Macau SAR, China · School of Computer Science, Peking University, Beijing, China · Key Laboratory of High Confidence Software Technologies (Peking University), Ministry of Education, Beijing, China

Streaming 3D Reconstruction Bayesian Filtering 3D ViTs

Open paper AskPDF arXiv

Feed-forward transformer models have driven remarkable progress in 3D vision,yet their quadratic complexity renders them impractical for long video sequences. Streaming 3D reconstruction addresses this bottleneck by processing frames sequen?tially with constant memory. However, existing recurrent architectures inevitably suffer from progressive degradation over long sequences, as they fail to account for the intrinsic reliability of observations and the accumulated confidence of the historical state, leading to severe error accumulation. To address these issues, we introduce BAT3R, a training-free Bayesian adaptation framework comprising two synergistic modules: State Innovation Gating (SIG) and Bayesian-guided Adaptive State Updating (BASU). SIG acts as a distributional regularizer that anchors state evolution to its geometric initialization, effectively arresting representational drift during long-term inference. Building on this, BASU reformulates state update as a recursive Bayesian inference process within a Kalman filtering framework. By jointly modeling the temporal reliability of observations via information-theoretic gain and spatial relevance via cross-attention maps, BASU achieves an asymmetric spatio-temporal state fusion: well-consolidated historical regions are protected from redundant updates, while genuinely novel scene content is aggressively incor?porated. Extensive experiments on standard benchmarks (7-Scenes, TUM, ScanNet, KITTI, etc.) demonstrate that BAT3R consistently outperforms state-of-the-art online 3D reconstruction methods, with the most significant gains observed on long sequences and complex scenes.

Abstract

Explore similar work

BAT3R: Robust Online 3D Reconstruction with Bayesian Adaptive State Updates

TTSA3R: Training-Free Temporal-Spatial Adaptive Persistent State for Streaming 3D Reconstruction

PAS3R: Pose-Adaptive Streaming 3D Reconstruction for Long Video Sequences