Click3R: Interactive Stereo 3D Reconstruction with Sparse Correspondence Clicks

Hao Gu; Jin Yao; Zezhou Cheng

cs.CVNov 9, 2026

Click3R: Interactive Stereo 3D Reconstruction with Sparse Correspondence Clicks

Authors: Hao Gu, Jin Yao, Zezhou Cheng

Organizations: University of Virginia

Abstract

Despite impressive progress, existing feed-forward large 3D reconstruction models produce a single, irrevocable prediction. When these models fail, there is no mechanism to correct their mistakes. To address this limitation, we introduce Click3R, a framework for interactive geometric correction of feed-forward models using sparse human correspondence clicks. Given an image pair and a small set of cross-view point correspondences provided by a user, Click3R injects geometric constraints into the 3D reconstruction networks through a lightweight point correspondence adapter, enabling the model to resolve ambiguities in challenging scenarios such as repetitive textures, symmetric structures, large viewpoint changes, or textureless regions. To evaluate interactive reconstruction, we curate a challenging benchmark specifically designed to expose failure modes of existing feed-forward 3D reconstruction methods. Experiments show that Click3R reduces reconstruction error with as few as a single correspondence click, improving performance on both standard and our curated benchmarks. Our results highlight that sparse human interaction provides an effective and practical mechanism for correcting geometric reconstruction errors.

Figures & tables

Figure 1: Overview. Given (a) an image pair and sparse correspondences ( colored dot clicks ), Click3R conditions DUSt3R [ 44 ] to produce accurate 3D reconstruction. (b) DUSt3R fails on a scene with limited overlaps. (c) With as few as two clicks, Click3R recovers correct geometry.

Figure 2: Click3R architecture. A shared ViT encoder processes the two input images. Correspondence prompts are tokenized using our correspondence tokenizer ( Section 3.2 ). A symmetric point correspondence adapter, implemented as a lightweight transformer with self-attention and cross-attention ( Section 3.3 ), updates the decoder-input features for each view. The asymmetric decoder then interprets the updated embeddings and passes them to the original DUSt3R [ 44 ] heads to predict 3D pointmaps and confidence.

Figure 3: Hard3DBench. Statistics shown for the 2-view stereo subset of Hard3DBench. (a) Hard3DBench produces substantially larger pretrained DUSt3R rotation errors than the regular benchmark. (b) Failure category distribution. (c) Example image pairs in Hard3DBench.

	CO3D		ScanNet++		ARKitScenes
Method	mAA@30° $\uparrow$	PM Err. $\downarrow$	mAA@30° $\uparrow$	PM Err. $\downarrow$	mAA@30° $\uparrow$	PM Err. $\downarrow$
VGGT [ 42 ] (518)	87.76	0.498	59.47	0.261	77.91	0.159
Pi3X [ 45 ] (518)	87.95	0.467	80.77	0.140	82.16	0.146
VGGT [ 42 ] (224)	62.53	1.994	24.47	1.325	40.89	0.690
Pi3X [ 45 ] (224)	85.02	0.577	72.53	0.196	76.39	0.167
DUSt3R [ 44 ]	79.20	0.728	82.57	0.190	86.02	0.146

Table 1: Standard benchmark results on held-out splits of CO3D, ScanNet++, and ARKitScenes. All methods use

224\!\times\!224

inputs; gray rows are VGGT and Pi3X at both 518 (native) and 224 resolutions, shown only as reference and not included in the bold/underline ranking. DUSt3R (retrained) is our re-trained DUSt3R baseline using the same training data as Click3R for fair comparison;

\pm

x.x denotes the mAA@30 difference from this baseline. PM Err.: median pointmap error. Best in bold , second best underlined .

In-Distribution (404 pairs / 400 tuples)
	2-view (pair)		5-view (tuple)
Method	mAA@30° $\uparrow$	Pointmap Err. $\downarrow$	mAA@30° $\uparrow$	Pointmap Err. $\downarrow$
VGGT [ 42 ] (518)	63.7	0.256	56.5	1.222
Pi3 [ 45 ] (518)	64.5	0.232	60.3	0.829
VGGT [ 42 ] (224)	40.5	0.620	33.2	1.536
Pi3 [ 45 ] (224)	60.7	0.299	56.9	1.234

Table 2: Hard3DBench results on the 2-view (left) and 5-view (right) subsets. In-distribution samples come from held-out splits of datasets used to train DUSt3R and Click3R ; out-of-distribution samples come from datasets absent from training. Click3R K denotes

K

GT correspondences as prompts (per pairwise forward in the 5-view setting). For 5-view evaluation, all DUSt3R-family methods (including Click3R ) follow DUSt3R’s global alignment protocol over the C(5,2) pairwise predictions. Comparison methods use

224\!\times\!224

resolution except MASt3R (512); gray rows are VGGT and Pi3 at both 518 (native) and 224 resolutions, shown only as reference and not included in the bold/underline ranking. Methods below the dashed line share identical training data; +x.x denotes mAA improvement over DUSt3R (retrained). Pointmap Error is the median across pairs/tuples. Best in bold , second best underlined .

Figure 4: Qualitative comparison. Each example features challenging geometry, such as large baselines, limited co-visibility, or substantial viewpoint changes. The input image pair with user-provided correspondence prompts ( colored dots ) is shown on the left, followed by each model’s reconstruction. Click3R produces more accurate and consistent 3D reconstructions, with prompted correspondences aligning more closely in 3D. Additional qualitative results are provided in Appendix Figure 13 .

Method	medRRE $\downarrow$	RRA@5° $\uparrow$	RRA@15° $\uparrow$	mAA@5 $\uparrow$	mAA@15 $\uparrow$	mAA@30 $\uparrow$
DUSt3R [ 44 ]	24.67°	1.0%	37.0%	0.0	9.5	25.6
DUSt3R (retrained)	15.05°	10.0%	50.0%	2.4	18.8	35.7
Click3R w/o prompts	17.30°	13.0%	47.0%	4.6	21.5	36.2
Click3R (human)	8.78°	22.0% +12.0	69.0% +19.0	9.2 +6.8	37.3 +18.5	56.2 +20.5
Click3R (random GT $K\!=\!4$ )	8.40°	27.0% +17.0	79.0% +29.0	9.4 +7.0	39.5 +20.7	59.2 +23.5

Table 3: Human-labeled correspondence evaluation. 100 out-of-distribution pairs with human-annotated correspondences. Click3R (human) uses sparse clicks from a human annotator; Click3R (random GT) uses

K\!=\!4

randomly sampled ground-truth correspondences as an oracle upper bound. +x.x denotes improvement over DUSt3R (retrained). Best in bold , second best underlined .

Panel of Figure 5 — Figure 5: Correspondence perturbation analysis. (a) Model confidence at the clicked pixels decreases as perturbation strength increases (CO3D, ScanNet++, and ARKitScenes). (b) 3D distance between the predicted points at the ground-truth correspondence (ARKitScenes): the model remains robust under small perturbations, but larger perturbations degrade matching performance.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Configuration	Params (M)	GPU time (ms)	Overhead
DUSt3R [ 44 ] (pretrained)	532.3	20.4	—
Click3R (no prompts)	566.4	20.4	$+$ 0%
Click3R ( $K\!=\!8$ )	566.4	21.5	$+$ 5%
Click3R ( $K\!=\!12$ )	566.4	21.4	$+$ 5%

Appendix

Table 4: Runtime and model complexity. GPU compute time measured on a single A100 (

224\!\times\!224

, bf16). Click3R without prompts has identical cost to DUSt3R;

K\!=\!8

prompts add 5% overhead.

Category	Detection criterion
Large viewpoint	Relative rotation $>60^{\circ}$
Moderate viewpoint	$30^{\circ}<$ relative rotation $\leq 60^{\circ}$
Low overlap	Co-visible pixel fraction $<20\%$
Large baseline	Baseline / median depth $>0.3$
Extreme depth range	$d_{\mathrm{max}}/d_{\mathrm{min}}>50$

Appendix

Table 5: Failure taxonomy for Hard3DBench. Categories are multi-label, detected from ground-truth poses and depth.

Dataset	Split	$N_{\text{pool}}$	$N_{\text{hard}}$	$N_{\text{final}}$	Scene type
In-Distribution (404 pairs, uniformly sampled from 3,406 candidates)
BlendedMVS [ 56 ]	val	19,719	1,000	97	Synthetic multi-view
CO3D [ 28 ]	test	19,396	1,000	109	Object-centric video
Habitat [ 32 ]	val	9,871	406	100	Synthetic indoor
WildRGB-D [ 50 ]	test	19,860	1,000	98	RGB-D video
Out-of-Distribution (300 pairs, uniformly sampled from 2,900 candidates)

Appendix

Table 6: Dataset composition of the 2-view subset of Hard3DBench.

N_{\text{pool}}

: pairs after overlap filtering;

N_{\text{hard}}

: candidates selected by worst rotation error;

N_{\text{final}}

: pairs after uniform subsampling. The 5-view subset uses the same datasets; per-dataset tuple counts are uniform (top-100 in-distribution, top-60 out-of-distribution).

Method	Lg.View.	Mod.View.	Low Ovlp.	Lg.Base.	Ext.Depth	All
DUSt3R [ 44 ]	35.9	52.0	46.3	44.6	38.9	41.8
DUSt3R (retrained)	54.3	56.3	48.7	58.0	41.3	54.6
Click3R w/o prompts	53.2	56.3	46.5	56.6	39.4	53.1
Click3R (8 prompts)	66.4 +12.1	60.6 +4.3	54.9 +6.2	66.0 +8.1	48.0 +6.6	61.7 +7.1

Appendix

Table 7: Per-category results on the 2-view subset of Hard3DBench. mAA@30° (

\uparrow

) grouped by failure taxonomy. Categories are multi-label (pairs may belong to multiple categories). Click3R with

K\!=\!8

GT correspondences. +x.x denotes improvement over DUSt3R (retrained).

No clicks
Training	CO3D	ScanNet++	ARKitScenes	H3B ID	H3B OOD	Mean
Adapter on frozen DUSt3R	79.2	82.6	86.0	39.2	47.1	66.8
Finetuned from pretrained DUSt3R	83.4	88.2	79.8	55.5	49.7	71.3
Trained from scratch (ours)	84.4	89.9	84.3	55.3	50.8	72.9
One click
Adapter on frozen DUSt3R	84.0	90.1	84.8	49.7	47.2	71.2

Appendix

Table 8: Ablation of training strategies on all evaluation sets (mAA@30°, %). All variants train the correspondence adapter. Same random GT clicks for every model. H3B: Hard3DBench 2-view. Best in bold .

	CO3D			ScanNet++			ARKitScenes
Method	@5°	@15°	@30°	@5°	@15°	@30°	@5°	@15°	@30°
VGGT [ 42 ] (518)	55.67	79.22	87.76	21.80	44.03	59.47	34.60	63.85	77.91
Pi3X [ 45 ] (518)	54.17	79.26	87.95	47.80	69.68	80.77	40.64	69.55	82.16
VGGT [ 42 ] (224)	4.37	36.78	62.53	0.20	7.14	24.47	1.26	17.37	40.89
Pi3X [ 45 ] (224)	44.27	73.85	85.02	31.80	57.70	72.53	31.68	61.41	76.39
DUSt3R [ 44 ]	33.25	65.17	79.20	50.24	73.83	82.57	48.46	75.15	86.02

Appendix

Table 9: Pose accuracy at stricter thresholds on the standard benchmark. mAA at 5°, 15° and 30° for every row of Table 1 , on the same pairs; the mAA@30° columns equal Table 1 . Gray rows are reference only and not ranked.

\pm

x.x denotes the difference from DUSt3R (retrained). Best in bold , second best underlined .

In-Distribution (404 pairs / 400 tuples)
	2-view (pair)			5-view (tuple)
Method	@5°	@15°	@30°	@5°	@15°	@30°
VGGT [ 42 ] (518)	27.4	50.1	63.7	32.5	48.5	56.5
Pi3 [ 45 ] (518)	34.3	53.1	64.5	41.0	54.4	60.3
VGGT [ 42 ] (224)	3.5	19.1	40.5	2.0	16.9	33.2
Pi3 [ 45 ] (224)	22.9	46.7	60.7	30.1	48.7	56.9

Appendix

Table 10: Pose accuracy at stricter thresholds on Hard3DBench. mAA at 5°, 15° and 30° for every row of Table 2 , over finite poses; 5-view values pool the ten pairwise poses of each tuple. The mAA@30° columns equal Table 2 . Gray rows are reference only and not ranked.

\pm

x.x denotes the difference from DUSt3R (retrained). Best in bold , second best underlined .

Figure 6: Prompt count ablation on Hard3DBench. Pose accuracy of Click3R as a function of

K

(red), with baselines shown as horizontal lines. Top: in-distribution (404 pairs); bottom: out-of-distribution (300 pairs). All methods use

224\!\times\!224

resolution.

Figure 7: Robustness to photometric augmentations. We perturb view 2 with brightness scaling or Gaussian noise and report RRA@15° on 100 hard OOD pairs. Prompts consistently improve accuracy, with the largest margin under heavy noise (

\sigma{=}0.3

): 55% vs. 6%.

Correct clicks (of 8)	medRRE $\downarrow$	mAA@30 $\uparrow$
100%	7.70°	62.8
75%	8.81°	55.6
50%	12.08°	44.9
25%	20.63°	30.0
0%	34.76°	17.7
No prompts	18.64°	35.6

Appendix

Table 11: Robustness to incorrect prompts. On the 100 pairs of Table 3 , we use 8 random ground-truth clicks and replace a fraction of the view-2 clicks by locations sampled uniformly over the image, at least 16 pixels from the true correspondence. Results are pooled over 5 random draws per pair.

	H3B ID		H3B OOD		H3B OOD subset
Clicks	RoMa v2	GT	RoMa v2	GT	RoMa v2	GT
$K\!=\!0$	55.9		49.4		35.6
$K\!=\!1$	58.4	57.8	57.0	55.8	48.7	48.1
$K\!=\!4$	58.0	60.5	59.6	60.8	55.0	58.7
$K\!=\!8$	58.4	62.4	59.4	61.8	58.5	61.1
Human	–		–		56.8

Appendix

Table 12: Automatic prompts from RoMa v2 [ 7 ] . mAA@30° of Click3R with clicks from RoMa v2 matches or from random ground-truth correspondences (GT). H3B: Hard3DBench 2-view (404 ID, 300 OOD pairs); the subset is the 100 OOD pairs with human clicks.

	H3B ID	H3B OOD
MASt3R (retrained)	61.3	50.6
MASt3R + adapter, $K\!=\!0$	61.1	52.1
MASt3R + adapter, $K\!=\!8$	62.4	54.9

Appendix

Table 13: Correspondence prompting on MASt3R. mAA@30° on Hard3DBench 2-view at 512 resolution (404 ID, 300 OOD pairs).

K\!=\!8

uses random ground-truth clicks.

Figure 8: Reconstruction loss distribution (aggregated over the three datasets of Table 1 ). Density estimated via KDE with Scott’s bandwidth. Click3R concentrates more mass near zero and exhibits a shorter tail than DUSt3R.

Figure 9: Pose error distribution (aggregated over the three datasets of Table 1 at

224\!\times\!224

\max(R_{\text{err}},T_{\text{err}})

in degrees). Click3R with prompts produces the sharpest peak near

0^{\circ}

and the shortest tail. The fraction of pairs exceeding

15^{\circ}

error drops from 9.4% (DUSt3R) to 2.8% ( Click3R with prompts), a

3.4{\times}

reduction in pose failures.

Figure 10: What the adapter changes. Per-token relative feature change induced by the adapter, with $K\!=\!8$ prompt locations shown as cyan dots. Bright regions indicate tokens whose features were modified most (each view uses its own color scale). The adapter updates every token, most strongly near the prompt correspondences.

Figure 11: How prompts reshape decoder cross-attention. Each row shows a query patch in view 1 (green cross) and its decoder cross-attention map over view 2. Without prompts (middle), attention is diffuse. With $K\!=\!8$ prompts (right, cyan dots), attention concentrates on the correct corresponding region.

Figure 12: A screenshot of the labeling tool we build to harness human labeled correspondences.

Figure 13: Additional qualitative comparisons. On challenging pairs, different baselines fail in different ways: DUSt3R may produce a distorted reconstruction on a pair where Pi3 succeeds, or vice versa. Click3R with sparse correspondence clicks consistently recovers accurate geometry.

Panel of Figure 14 — Figure 14: Failure cases of Click3R . Each panel: input pair (left), ground truth (middle), Click3R with clicks (right). (a) Near-zero overlap showcasing fundamental limitation of click-based supervision. (b) Extreme rotations cause geometric distortion that clicks cannot resolve. (c) In large-scale scenes, regions far from clicked correspondences still deform. (d) Clicks on featureless surfaces are inherently ambiguous and provide insufficient geometric signal.

Asset	License
Datasets (training)
CO3D [ 28 ]	CC-BY-NC 4.0
ARKitScenes [ 2 ]	Apple License (non-commercial primary)
ScanNet++ [ 57 ]	ScanNet++ Terms of Use
WildRGB-D [ 50 ]	MIT License
BlendedMVS [ 56 ]	CC-BY 4.0

Appendix

Table 14: Licenses of assets used.

3D Reconstruction

3D reconstruction recovers a scene’s shape, appearance, and camera motion from images, video, depth, or other sensor measurements.

See all 992 papers

Free Geometry: Refining 3D Reconstruction from Longer Versions of Itself
Apr 15, 2026Yuhang Dai, Xingyi YangFeed-Forward 3D Reconstruction Multi-View 3D Reconstruction
The Hong Kong Polytechnic University, Hong Kong SAR
Any3R: Feedforward 3D Reconstruction from Any Camera Type
Oct 26, 2026Jingong Chen, Qingwen Zhang, Youming Deng +53D Reconstruction Multi-View 3D Reconstruction
South Dakota State University · Aalto University · KTH Royal Institute of Technology +3
CAM3R: Camera-Agnostic Model for 3D Reconstruction
Mar 23, 2026Namitha Guruprasad, Abhay Yadav, Cheng Peng +13D Reconstruction Multi-View 3D Reconstruction
Johns Hopkins University, Baltimore, MD 21218, USA · University of Virginia, Charlottesville, VA 22904, USA

Apr 15, 2026J/K move · Enter open · S save

Free Geometry: Refining 3D Reconstruction from Longer Versions of Itself

Yuhang Dai, Xingyi Yang

The Hong Kong Polytechnic University, Hong Kong SAR

Feed-Forward 3D Reconstruction Multi-View 3D Reconstruction Test-Time Adaptation Self-Supervised Learning

Open paper AskPDF arXiv

Feed-forward 3D reconstruction models are efficient but rigid: once trained, they perform inference in a zero-shot manner and cannot adapt to the test scene. As a result, visually plausible reconstructions often contain errors, particularly under occlusions, specularities, and ambiguous cues. To address this, we introduce Free Geometry, a framework that enables feed-forward 3D reconstruction models to self-evolve at test time without any 3D ground truth. Our key insight is that, when the model receives more views, it produces more reliable and view-consistent reconstructions. Leveraging this property, given a testing sequence, we mask a subset of frames to construct a self-supervised task. Free Geometry enforces cross-view feature consistency between representations from full and partial observations, while maintaining the pairwise relations implied by the held-out frames. This self-supervision allows for fast recalibration via lightweight LoRA updates, taking less than 2 minutes per dataset on a single GPU. Our approach consistently improves state-of-the-art foundation models, including Depth Anything 3 and VGGT, across 4 benchmark datasets, yielding an average improvement of 3.73% in camera pose accuracy and 2.88% in point map prediction. Code is available at https://github.com/hiteacherIamhumble/Free-Geometry .

Abstract

3D Reconstruction

Explore similar work

Free Geometry: Refining 3D Reconstruction from Longer Versions of Itself

Any3R: Feedforward 3D Reconstruction from Any Camera Type

CAM3R: Camera-Agnostic Model for 3D Reconstruction