Graph3R: Unifying View Graph Initialization with Feed-Forward Reconstruction

Jingwen Yu; Hengli Wang; Jianhao Jiao; Zhonghang Liu; Ping Tan; Hong Zhang

cs.CVOct 12, 2026

Graph3R: Unifying View Graph Initialization with Feed-Forward Reconstruction

Authors: Jingwen Yu, Hengli Wang, Jianhao Jiao, Zhonghang Liu, Ping Tan, Hong Zhang

Organizations: Hong Kong University of Science and Technology · Southern University of Science and Technology · China Merchants Group, LionRock AI Lab · SmartMore Ltd.

Abstract

Feed-forward reconstruction recovers camera poses and dense geometry from image sets whose views are assumed to come from a single scene. In practice, uncurated image collections often contain views from disjoint scenes, and reconstructing these views jointly can corrupt camera poses and geometry. Existing pipelines typically resolve view associations in separate stages before final reconstruction, leaving view graph initialization and reconstruction inference only loosely coupled. We introduce Graph3R, which unifies view graph initialization with feed-forward reconstruction through a shared geometric representation. A lightweight verification head predicts pairwise covisibility from the multi-view features of $π^3$ to construct a view graph. This graph classifies the input views into scene groups. For each group, we mask the shared features before decoding camera poses and dense geometry. Reusing the multi-view features allows Graph3R to reconstruct disjoint scenes separately with a single backbone pass. Across three benchmarks, Graph3R improves average maximum recall at 100% precision by upto 8.99 percentage points over the baseline, while improving reconstruction quality on mixed-scene inputs. Regarding the multi-session mapping task, Graph3R improves cross-session localization and achieves competitive pose estimation accuracy, while delivering a runtime speedup of approximately $35.9\times$ over the baseline when reconstructing the query group. By integrating learned view association with geometric inference, our approach takes a step toward end-to-end reconstruction from uncurated image collections.

Figures & tables

Figure 1: Comparison with existing baselines , given input views from three disjoint scenes (A–C), the baseline method ${\pi}^{3}$ ( Wang et al., 2025b ) produces an incorrect reconstruction (left). While ours ( Graph3R ) predicts a view graph from shared geometric features, partitions the views into scene groups, and reconstructs each group separately using a single backbone pass (right).

Figure 2: Comparison of different SfM paradigms. Classical SfM Pan et al. (2024) and GLUEMAP ( Pan et al., 2026 ) initialize a view graph through separate verification stages before reconstruction. RobustVGGT ( Han et al., 2025 ) selects views using VGGT’s latent features and reruns the feed forward model on the selected views. Graph3R proposes a unified paradigm that jointly predicts the view graph and scene structure with a single backbone inference as detailed illustrated in Fig. 3 .

Figure 3: System overview of Graph3R . (a) The $\pi^{3}$ baseline ( Wang et al., 2025b ) . (b) Graph3R introduces a verification head that predicts pairwise covisibility before geometry decoding. These predictions are used to construct a view graph, which guides the graph-conditioned reconstruction as described in Sec. 3.3 , enabling each scene to be reconstructed individually.

Figure 4: Architecture of the verification head. The head takes the shared feature representation $\mathbf{Z}$ as input and predicts pairwise covisibility scores $\mathbf{S}$ . The scores are used to construct a view graph $\mathcal{G}$ , which defines the mask of each scene group. The masks are applied to the shared representation $\mathbf{Z}$ prior to the geometry decoder, enabling graph-conditioned reconstruction for each scene group.

$K$	Individual		Graph3R		$\Delta$ (%)
$K$	AbsRel $\downarrow$	$\delta_{1}$ (%) $\uparrow$	AbsRel $\downarrow$	$\delta_{1}$ (%) $\uparrow$	AbsRel $\downarrow$	$\delta_{1}$ (%) $\uparrow$
2	0.098	92.9	0.098	91.9	0.00	$-1.08$
3	0.097	93.4	0.097	92.1	0.00	$-1.39$
4	0.098	93.2	0.102	91.4	$+4.08$	$-1.93$

Table 1: Controlled feature reuse on ETH3D. We separately report AbsRel

\downarrow

and

\delta_{1}

(%)

\uparrow

for independent decoding and group decoding from one shared backbone representation.

\Delta

reports relative percentage changes in each metric in comparison to the Individual baseline.

Method	GSV-Cities		360Loc		ETH3D		Average
Method	mAP $\uparrow$	MR $\uparrow$	mAP $\uparrow$	MR $\uparrow$	mAP $\uparrow$	MR $\uparrow$	mAP $\uparrow$	MR $\uparrow$
GLUEMAP † ( Pan et al., 2026 )	91.17	27.29	96.13	74.79	96.86	93.17	94.72	65.08
RobustVGGT ( Han et al., 2025 )	70.85	3.46	91.89	15.36	86.58	49.97	83.11	22.93
RobustPi3 ( Han et al., 2025 )	74.89	0.20	93.46	6.31	97.76	65.38	88.70	23.96
Graph3R (Ours)	98.98	47.17	99.65	83.64	99.38	91.41	99.34	74.07

Table 2: Benchmark of View Graph Initialization.

Data Mixture				Contr. Loss Weight			Trainable Modules
Backbone	Data	AP $\uparrow$	MR $\uparrow$	$\lambda$	AP $\uparrow$	MR $\uparrow$	Backbone	Params.	AP $\uparrow$	MR $\uparrow$
Frozen	GSV	93.75	37.32	0	99.14	57.25	Frozen	69M	98.32	57.07
Frozen	4:1	98.11	57.46	0.1	99.22	64.29	$\hookrightarrow$ FT2	103M	97.97	53.44
LoRA	GSV	97.67	53.07	0.25	99.31	73.06	$\hookrightarrow$ FT4	128M	98.70	38.49
LoRA	4:1	99.14	57.25	0.5	99.09	54.82	$\hookrightarrow$ LoRA	80M	99.31	73.06

Table 3: Ablations of the verification head. All the Data Mixture uses

\lambda=0

, Contrastive Loss Weight and Trainable Modules use the 4:1 mixture, with LoRA and

\lambda=0.25

fixed, respectively. Params. counts include the verification head.

Method	Easy			Hard			Average
Method	MVS $\uparrow$	AbsRel $\downarrow$	$\delta_{1}$ (%) $\uparrow$	MVS $\uparrow$	AbsRel $\downarrow$	$\delta_{1}$ (%) $\uparrow$	MVS $\uparrow$	AbsRel $\downarrow$	$\delta_{1}$ (%) $\uparrow$
RobustPi3 ( Han et al., 2025 )	0.230	0.287	73.6	0.133	0.394	65.9	0.1815	0.3405	69.75
GLUEMAP ( Pan et al., 2026 )	0.127	0.549	61.2	0.078	0.586	59.9	0.1025	0.5675	60.55
Graph3R (ours)	0.271	0.264	77.5	0.175	0.321	70.9	0.2230	0.2925	74.20

Table 4: Multi-view depth consistency on ETH3D ( Schops et al., 2017 ) .

Method	ETH3D		OpenNavMap		360Loc
Method	$10^{\circ}$ /1m	$20^{\circ}$ /5m	$10^{\circ}$ /1m	$20^{\circ}$ /5m	$10^{\circ}$ /1m	$20^{\circ}$ /5m
Pi3 ( Wang et al., 2025b )	27.55	41.04	19.87	31.17	1.86	16.86
VGGT ( Wang et al., 2025a )	8.43	22.99	8.89	28.60	0.25	12.06
RobustPi3 ( Han et al., 2025 )	73.08	85.71	60.92	68.36	54.07	64.71
RobustVGGT ( Han et al., 2025 )	54.90	62.82	48.13	70.63	49.75	62.75
GLUEMAP ( Pan et al., 2026 )	60.00	72.04	67.83	79.72	52.79	66.37

Table 5: Benchmark of pose estimation accuracy. We report the accuracy at different thresholds.

Method	Region A				Region B
Method	R@1	R@3	R@5	R@8	R@1	R@3	R@5	R@8
CosPlace ( Berton et al., 2022 )	63.93	69.48	73.01	76.04	66.05	71.60	73.95	76.49
RobustPi3 ( Han et al., 2025 )	65.83	73.77	75.28	–	70.35	74.30	75.56	–
RobustVGGT ( Han et al., 2025 )	71.25	74.65	75.28	–	71.35	74.21	75.26	–
GLUEMAP † ( Pan et al., 2026 )	72.26	75.41	75.91	–	71.33	73.74	75.19	–
Graph3R(Ours)	74.15	75.54	76.04	–	72.88	75.00	75.79	–

Table 6: Cross-session localization on OpenNavMap. All methods rerank on CosPlace retrieved candidates.

Figure 5: Efficiency comparison with Doppelgangers++ ( Xiangli et al., 2025 ) . Our method achieves approximately 245

\times

speedup as the number of input images

N=16

Method	All groups		Query group
Method	s/batch $\downarrow$	Peak Memory $\downarrow$	s/batch $\downarrow$	Peak Memory $\downarrow$
RobustPi3 ( Han et al., 2025 )	0.218	4.727	0.158	4.727
RobustVGGT ( Han et al., 2025 )	0.693	11.065	0.421	11.065
GLUEMAP ( Pan et al., 2026 )	4.663	4.635	4.564	4.637
Graph3R (Ours)	0.159	8.087	0.127	8.087

Table 7: Computational efficiency on OpenNavMap. End-to-end latency and peak GPU memory (GB) for

N=9

images per batch, decoding all predicted groups or only the group containing the query.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Feature context	Association	Reconstruction
MegaLoc	Per-image	Similarity	External stage
RoMAv2 + RANSAC	Pairwise	Geometric verification	External stage
DG / DG++	Pairwise	Learned classifier	SfM
RobustVGGT	Multi-view	Attention / features	Backbone rerun
GLUEMAP	Image + pair	Retrieval + verification	Local inference + global optimization
Graph3R (ours)	Multi-view	Learned classifier	Group decoding from retained features

Appendix

Table 8: View association and subsequent reconstruction. DG denotes Doppelgangers; feature context refers to association.

Adaptation	Data	$\lambda$	GSV-Cities				360Loc
Adaptation	Data	$\lambda$	Easy	Medium	Hard	Full	360Loc
Frozen	1:0	0.0	99.45/67.62	99.55/58.38	98.38/39.16	98.59/19.58	88.51/28.46
Frozen	4:1	0.0	99.12/58.53	98.10/44.05	94.78/23.66	97.00/19.93	98.96/78.37
LoRA	1:0	0.0	99.83/84.86	99.69/90.41	98.92/66.67	99.14/54.01	95.94/32.15
LoRA	4:1	0.0	99.17/41.04	99.10/42.44	97.93/17.81	98.69/18.56	99.56/84.54
LoRA	4:1	0.1	99.08/75.44	99.13/77.92	98.07/8.48	98.70/8.33	99.69/86.04

Appendix

Table 10: Per-split results for the main ablations. Each metric entry is AP/MR@100P (%). Data is the GSV-Cities:360Loc ratio;

\lambda=0

disables the contrastive term. The main-table mean weights the mean over four GSV-Cities splits and 360Loc equally. Configurations shared by multiple ablations appear only once.

Grouping scorer	Easy			Hard			Average
Grouping scorer	MVS $\uparrow$	AbsRel $\downarrow$	$\delta_{1}$ (%) $\uparrow$	MVS $\uparrow$	AbsRel $\downarrow$	$\delta_{1}$ (%) $\uparrow$	MVS $\uparrow$	AbsRel $\downarrow$	$\delta_{1}$ (%) $\uparrow$
RobustPi3 ( Han et al., 2025 )	0.230	0.287	73.6	0.133	0.394	65.9	0.1815	0.3405	69.75
RobustVGGT	0.290	0.239	78.8	0.157	0.393	71.0	0.2235	0.3160	74.90
GLUEMAP ( Pan et al., 2026 )	0.127	0.549	61.2	0.078	0.586	59.9	0.1025	0.5675	60.55
Graph3R (ours)	0.271	0.264	77.5	0.175	0.321	70.9	0.2230	0.2925	74.20

Appendix

Table 11: ETH3D depth consistency benchmark. Average is the equal-weight mean of the displayed Easy and Hard results. Bold and underline mark the best and second-best results in each column, respectively.

(a) Easy (clique)
$K$	Grouping scorer	View graph initialization (%)			Depth consistency
$K$	Grouping scorer	AP $\uparrow$	MR $\uparrow$	F1 $\uparrow$	MVS $\uparrow$	AbsRel $\downarrow$	$\delta_{1}$ (%) $\uparrow$
2	RobustPi3	99.74	95.42	93.57	0.350	0.189	83.1
	RobustVGGT	99.86	89.58	70.62	0.388	0.181	84.2
	GLUEMAP	97.99	93.75	79.86	0.216	0.360	72.4
	Graph3R (ours)	99.92	99.17	98.09	0.339	0.191	82.8

Appendix

Table 12: View graph initialization and multi-view depth consistency on ETH3D. Each setting contains 20 fixed episodes with

K

scenes and four views per scene. Easy (clique) requires pairwise within-scene overlap; Hard (connected) requires a connected overlap graph. Only the grouping scorer varies: reconstruction uses the same

\pi^{3}

weights, shared features, and grouped decoding at

224\!\times\!224

ETH3D
Method	AP-RPE median $\downarrow$		SA-RPE mean $\downarrow$		System accuracy (%) $\uparrow$
Method	Rot. ( ∘ )	Dir. ( ∘ )	Rot. ( ∘ )	Dir. ( ∘ )	$10^{\circ}$ /1 m	$20^{\circ}$ /5 m
Pi3 ( Wang et al., 2025b )	4.67	16.04	99.60	107.58	27.55	41.04
VGGT ( Wang et al., 2025a )	14.76	43.25	117.39	126.79	8.43	22.99
RobustPi3 ( Han et al., 2025 )	4.46	15.52	20.70	28.53	73.08	85.71
RobustVGGT ( Han et al., 2025 )	8.57	25.90	61.44	65.45	54.90	62.82

Appendix

Table 13: Full pose estimation accuracy. Supplement to Table 5 , using the same strict score

>0.5

verification rule. AP-RPE and SA-RPE are defined in Appendix A.6 . Bold and underline mark the best and second-best results per dataset and metric.

	Region A				Region B				OpenNavMap (macro)
Ranking	R@1	R@3	R@5	R@8	R@1	R@3	R@5	R@8	R@1	R@3	R@5	R@8
CosPlace	63.93	69.48	73.01	76.04	66.05	71.60	73.95	76.49	64.99	70.54	73.48	76.26
RobustPi3	65.83	73.77	75.28	–	70.35	74.30	75.56	–	68.09	74.04	75.42	–
RobustVGGT	71.25	74.65	75.28	–	71.35	74.21	75.26	–	71.30	74.43	75.27	–
GLUEMAP †	72.26	75.41	75.91	–	71.33	73.74	75.19	–	71.79	74.58	75.55	–
Graph3R (Ours)	74.15	75.54	76.04	–	72.88	75.00	75.79	–	73.52	75.27	75.92	–

Appendix

Table 14: Cross-session localization on OpenNavMap. CosPlace retrieves the top eight candidates per query, which each method reranks using its saved verification scores. Recall@K is the fraction of answerable queries with at least one correct result among the top K; correctness uses ground-truth camera-center distance

\leq 25

m and relative orientation

\leq 40^{\circ}

. All entries are percentages. The CosPlace R@8 columns show the retrieval ceiling shared by every reranker. OpenNavMap is the equal-weight mean of Region A and Region B.

Method	s/batch $\downarrow$	batch/s $\uparrow$	Equivalent pairs/s $\uparrow$	Avg peak GB $\downarrow$	Max peak GB $\downarrow$
RobustPi3	0.218	4.588	165.180	4.727	4.727
RobustVGGT	0.693	1.442	51.912	10.994	11.065
GLUEMAP	4.663	0.214	7.721	4.633	4.635
Graph3R (Ours)	0.159	6.305	226.995	8.087	8.087

Appendix

Table 15: OpenNavMap runtime and peak GPU memory: all predicted groups. Each batch contains nine images and 36 unordered pairs. Each method uses its own verification scores with the strict score

>0.5

connected-component rule. Equivalent pairs/s is 36 times batch/s. Average and maximum peak memory subtract each run’s measured idle baseline and use decimal GB. Best and second-best results are bold and underlined, respectively.

Method	s/batch $\downarrow$	batch/s $\uparrow$	Equivalent pairs/s $\uparrow$	Avg peak GB $\downarrow$	Max peak GB $\downarrow$
RobustPi3	0.158	6.327	227.773	4.727	4.727
RobustVGGT	0.421	2.376	85.550	10.995	11.065
GLUEMAP	4.564	0.219	7.887	4.634	4.637
Graph3R (Ours)	0.127	7.851	282.621	8.087	8.087

Appendix

Table 16: OpenNavMap runtime and peak GPU memory: query group only. The output scope is restricted to the connected component containing query image 0, using the same nine-image, 36-pair batches and strict score

>0.5

rule as Table 15 . Equivalent pairs/s is 36 times batch/s. Average and maximum peak memory subtract each run’s measured idle baseline and use decimal GB. Best and second-best results are bold and underlined, respectively.

Method	VisymScenes		GV-Bench		Avg.
Method	AP $\uparrow$	MR $\uparrow$	AP $\uparrow$	MR $\uparrow$	AP $\uparrow$	MR $\uparrow$
MegaLoc ( Berton and Masone, 2025 )	91.17	26.42	91.74	32.38	91.46	29.40
RoMAv2 ( Edstedt et al., 2025 )	92.75	2.45	84.63	19.49	88.69	10.97
RobustVGGT ( Han et al., 2025 )	93.57	18.81	52.13	2.14	72.85	10.48
Doppelgangers ( Cai et al., 2023 )	97.89	38.21	73.95	17.28	85.92	27.75
Doppelgangers++ ( Xiangli et al., 2025 )	98.42	41.76	89.70	25.75	94.06	33.76

Appendix

Table 17: Pairwise verification. We evaluate on VisymScenes ( Xiangli et al., 2025 ) and GV-Bench ( Yu et al., 2024 ) with average precision (AP) and maximum recall at 100% precision (MR), both expressed as percentages. Avg. reports the mean of each metric across the two benchmarks. The first three methods are not trained on VisymScenes; the remaining methods are. Bold marks the best result per column.

Method	Day		Night		Weather		Season		Nordland		Uacampus		Avg.
Method	AP $\uparrow$	MR $\uparrow$	AP $\uparrow$	MR $\uparrow$	AP $\uparrow$	MR $\uparrow$	AP $\uparrow$	MR $\uparrow$	AP $\uparrow$	MR $\uparrow$	AP $\uparrow$	MR $\uparrow$	AP $\uparrow$	MR $\uparrow$
MegaLoc	99.64	57.84	97.81	28.58	99.77	49.26	99.81	48.72	77.32	3.15	76.09	6.72	91.74	32.38
RoMAv2	99.78	50.31	97.89	13.47	99.82	14.89	99.75	7.33	29.30	0.11	81.22	30.85	84.63	19.49
Graph3R (FT last 2)	98.53	32.39	80.00	1.63	99.57	13.68	99.34	16.20	46.21	0.17	65.06	2.99	81.45	11.18
Graph3R (LoRA)	99.20	40.63	87.15	3.79	99.83	18.90	99.46	27.90	52.19	1.30	73.17	2.24	85.17	15.79
RobustVGGT	78.74	7.66	36.19	0.19	94.93	1.08	88.76	0.06	6.42	0.11	7.76	3.73	52.13	2.14

Appendix

Table 18: Pairwise verification on GV-Bench (per condition). We report AP and MR (%) across the six GV-Bench conditions; Avg. is the mean over conditions. Best result per column in bold .

Figure 6: Visualization of multi-session mapping on OpenNavMap. Three cases are shown, each with cross-session associations above and the corresponding mapping results below.

Explore similar work

Apr 10, 2026

FF3R: Feedforward Feature 3D Reconstruction from Unconstrained views

Recent advances in vision foundation models have revolutionized geometry reconstruction and semantic understanding. Yet, most of the existing approaches treat these capabilities in isolation, leading to redundant pipelines and compounded errors. This paper introduces FF3R, a fully annotation-free feed-forward framework that unifies geometric and semantic reasoning from unconstrained multi-view image sequences. Unlike previous methods, FF3R does not require camera poses, depth maps, or semantic labels, relying solely on rendering supervision for RGB and feature maps, establishing a scalable paradigm for unified 3D reasoning. In addition, we address two critical challenges in feedforward feature reconstruction pipelines, namely global semantic inconsistency and local structural inconsistency, through two key innovations: (i) a Token-wise Fusion Module that enriches geometry tokens with semantic context via cross-attention, and (ii) a Semantic-Geometry Mutual Boosting mechanism combining geometry-guided feature warping for global consistency with semantic-aware voxelization for local coherence. Extensive experiments on ScanNet and DL3DV-10K demonstrate FF3R's superior performance in novel-view synthesis, open-vocabulary semantic segmentation, and depth estimation, with strong generalization to in-the-wild scenarios, paving the way for embodied intelligence systems that demand both spatial and semantic understanding.

Chaoyi Zhou, Run Wang, Feng Luo +4

Microsoft · Clemson University · Texas A&M University

3D Reconstruction Feed-Forward 3D Reconstruction Multi-View 3D Reconstruction Open-Vocabulary 3D Segmentation

Ask PDF

Mar 24, 2026

UniQueR: Unified Query-based Feedforward 3D Reconstruction

We present UniQueR, a unified query-based feedforward framework for efficient and accurate 3D reconstruction from unposed images. Existing feedforward models such as DUSt3R, VGGT, and AnySplat typically predict per-pixel point maps or pixel-aligned Gaussians, which remain fundamentally 2.5D and limited to visible surfaces. In contrast, UniQueR formulates reconstruction as a sparse 3D query inference problem. Our model learns a compact set of 3D anchor points that act as explicit geometric queries, enabling the network to infer scene structure, including geometry in occluded regions--in a single forward pass. Each query encodes spatial and appearance priors directly in global 3D space (instead of per-frame camera space) and spawns a set of 3D Gaussians for differentiable rendering. By leveraging unified query interactions across multi-view features and a decoupled cross-attention design, UniQueR achieves strong geometric expressiveness while substantially reducing memory and computational cost. Experiments on Mip-NeRF 360 and VR-NeRF demonstrate that UniQueR surpasses state-of-the-art feedforward methods in both rendering quality and geometric accuracy, using an order of magnitude fewer primitives than dense alternatives.

Chensheng Peng, Quentin Herau, Jiezhi Yang +6

Applied Intuition · UC Berkeley · Stanford University

3D Gaussian Splatting 3D Reconstruction Multi-View 3D Reconstruction 3D Scene Representation Feed-Forward 3D Reconstruction

Ask PDF

Jul 10, 2026

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

Recent 3D geometric foundation models, such as VGGT, provide robust feed-forward 3D reconstruction by directly predicting camera poses and 3D scene points from input images. However, their results remain inaccurate, and scaling them to long sequences or large unordered image sets typically requires chunk-wise processing, which can introduce drift and inconsistency. We present Glob3R, a global SfM-style reconstruction built on 3D foundation models. Our key idea is to explicitly optimize feed-forward geometric predictions. To this end, we augment a frozen Pi3X backbone with a lightweight dense matching head that predicts image warps between selected reference frames and neighboring views. These dense warps are converted into sparse but reliable multi-view feature tracks, which provide correspondence constraints for global optimization. We further introduce a keyframe-based sliding-window association strategy that propagates tracks and relative poses across overlapping windows, enabling scalable reconstruction. Finally, we perform global motion averaging and bundle adjustment to refine camera poses, reduce scale inconsistencies, and recover dense scene geometry. Extensive experiments on indoor, outdoor, large-scale driving, and unordered SfM benchmarks demonstrate that Glob3R achieves robust and accurate reconstruction. It consistently improves over feed-forward foundation-model baselines and recent scalable reconstruction methods, while being more robust than classical SfM pipelines. The refined poses also lead to higher-quality neural rendering, validating the benefit of combining foundation-model priors with global geometric optimization. Project page: https://junyuandeng.github.io/Glob3r

Junyuan Deng, Heng Li, Kejie Qiu +7

The Hong Kong University of Science and Technology · Tongyi Lab, Alibaba Group · Nanjing University +1

Bundle Adjustment Multi-View 3D Reconstruction Structure from Motion

Ask PDF