Any3R: Feedforward 3D Reconstruction from Any Camera Type

Jingong Chen; Qingwen Zhang; Youming Deng; Chulwoo Pack; Olaf Wysocki; Boris Jutzi; Kwanghee Won; Kyle Gao

cs.CVOct 26, 2026

Any3R: Feedforward 3D Reconstruction from Any Camera Type

Authors: Jingong Chen, Qingwen Zhang, Youming Deng, Chulwoo Pack, Olaf Wysocki, Boris Jutzi, Kwanghee Won, Kyle Gao

Organizations: South Dakota State University · Aalto University · KTH Royal Institute of Technology · Cornell University ·

Abstract

Geometry foundation models recover camera poses and dense geometry from a set of images with a feed-forward network, but they expect pinhole images. That assumption is inside the backbone too: the positional encoding and the training loss both bind an image position to a direction in space, and both depend on the lens. We condition both on the dense ray field, the unit direction through every pixel, which describes any central camera without a lens type, distortion coefficients or a bound on field of view. Angular coordinates give the rotary encoding the measured ray angle of each patch, evaluated per view, and solid-angle weighting scales the point loss by the area each pixel covers on the sphere, so neither the encoding nor the supervision depends on the lens. Any3R is therefore a feed-forward model that takes any central camera, and is robust to mixing many camera types per scene, with no retraining per lens. On a benchmark of thirteen central cameras across three datasets, where every method reconstructs with its own predicted camera poses, it is the most consistently accurate method. Any3R enables scalable 3D reconstruction from heterogeneous crowdsourced and archival imagery collections sourced across diverse camera types. Project page: https://jingong-chen.github.io/Any3R/

Figures & tables

Figure 2: Motivation. Read from the pixel grid (ii), a patch position stands for a different direction and a pixel for a different solid angle on every lens (i), and the ray field (iii) corrects both. In the bottom row the ray field is withheld, so the backbone reads every frame as a panorama and the reconstruction follows that reading, which is correct only at $360^{\circ}$ .

Figure 3: Overview. The ray field $\bm{R}$ of each view gives its patches angular coordinates for the rotary encoding in intra-frame attention, and the heads predict the camera poses and a radial distance $d$ per pixel, decoded as $\bm{X}=d\bm{R}$ . In training, the point loss $\mathcal{L}_{\mathrm{pt}}$ compares $\bm{X}$ with the ground truth $\bm{X}^{\mathrm{gt}}$ , and the solid angle $\mathrm{d}\Omega$ of the same $\bm{R}$ weighs the error of each pixel, shown here on the panorama, where cells of equal pixel count cover unequal areas of the sphere.

Acc $\downarrow$ (m)
	Indoor								Outdoor
Method	Pin.	B–C	Equi.	ERP-C	MEI	KB4	DS	ERP	Pin.	B–C	Equi.	ERP-C	MEI	KB4	DS	ERP
	$90^{\circ}$	$90^{\circ}$	$180^{\circ}$	$180^{\circ}$	$180^{\circ}$	$200^{\circ}$	$200^{\circ}$	$360^{\circ}$	$90^{\circ}$	$90^{\circ}$	$180^{\circ}$	$180^{\circ}$	$180^{\circ}$	$200^{\circ}$	$200^{\circ}$	$360^{\circ}$
OmniV [ 39 ]	0.250	0.201	0.326	0.318	0.329	0.355	0.350	0.406	2.875	2.676	3.313	2.804	3.703	3.941	3.670	1.479
Wid3R [ 22 ]	0.144	–	0.141	–	0.146	–	–	0.101	2.278	–	2.121	–	2.252	–	–	2.330
PanoV [ 18 ]	0.282	0.272	0.259	0.348	0.265	0.251	0.265	0.072	2.256	2.708	2.498	3.633	2.433	2.405	2.473	0.751

Table 1: Comparison with the methods built for other cameras on the eight camera models used in training. Acc

\downarrow

(m) is the mean distance from a reconstructed point to its nearest ground-truth point, and AUC

\uparrow

(%) is the area under the curve of the pairwise pose error up to

15^{\circ}

and

30^{\circ}

; every method fuses its views with its own predicted poses. Indoor averages Matterport3D and Stanford2D3DS, and outdoor is PanoCity. The cameras are pinhole (Pin.), Brown–Conrady (B–C), equidistant (Equi.), cropped equirectangular (ERP-C), MEI, KB4, double sphere (DS) and equirectangular (ERP), with their fields of view below them. OmniV and PanoV abbreviate OmniVGGT and PanoVGGT, Wid3R receives the true ray field, MapA-ray is MapAnything with the true ray field fed to its ray encoder, and – marks cameras a method cannot take. In all tables, best and second best are shaded.

Figure 4: Qualitative comparison with the camera-aware methods. One camera per row, every method fused with its own predicted poses and rendered from one viewpoint per row. Cells marked n/a are structural, since Argus accepts only full panoramas and Wid3R’s camera classes cannot represent the $200^{\circ}$ KB4 lens. MapA-ray is MapAnything given the true ray field, and PaGeR, which predicts no pose, is not shown.

	ScanNet++			FIORD
Method	Acc $\downarrow$	@15 $\uparrow$	@30 $\uparrow$	Acc $\downarrow$	@15 $\uparrow$	@30 $\uparrow$
OmniV [ 39 ]	0.097	18.6	41.5	0.363	18.6	40.3
Wid3R † [ 22 ]	0.039	85.3	92.6	–	–	–
PanoV [ 18 ]	0.277	1.8	9.7	0.511	29.4	51.2
MapA-ray [ 25 ]	0.093	18.5	43.6	0.396	19.4	36.5
$\pi^{3}$ [ 61 ]	0.086	20.6	46.6	0.378	15.7	36.7

Table 2: Zero-shot reconstruction and camera pose on real fisheye lenses. ScanNet++ is its DSLR fisheye on its validation split and FIORD the

200^{\circ}

lens of its rig; neither is in Any3R’s training data. Measures and protocol as in Tab. 1 ; Any3R, MapA-ray and Wid3R receive the calibrated ray field, and – marks a lens a method cannot take. † Trained on ScanNet++. Any3R ∗ is not ranked.

Acc $\downarrow$ (m)
	Indoor								Outdoor
Method	Pin.	B–C	Equi.	ERP-C	MEI	KB4	DS	ERP	Pin.	B–C	Equi.	ERP-C	MEI	KB4	DS	ERP
	$90^{\circ}$	$90^{\circ}$	$180^{\circ}$	$180^{\circ}$	$180^{\circ}$	$200^{\circ}$	$200^{\circ}$	$360^{\circ}$	$90^{\circ}$	$90^{\circ}$	$180^{\circ}$	$180^{\circ}$	$180^{\circ}$	$200^{\circ}$	$200^{\circ}$	$360^{\circ}$
$\pi^{3}$ [ 61 ]	0.084	0.119	0.171	0.287	0.164	0.189	0.183	0.414	1.479	1.711	2.093	2.426	2.036	2.394	2.271	1.628
MapA [ 25 ]	0.092	0.149	0.273	0.269	0.267	0.283	0.294	0.386	1.485	1.633	2.066	2.861	2.236	3.297	2.789	1.833
DA3 [ 34 ]	0.313	0.294	0.361	0.288	0.363	0.396	0.379	0.402	4.360	5.689	6.656	2.717	6.844	6.463	6.477	3.149

Table 3: Comparison with geometry foundation models on the eight camera models used in training. Measures, domains and cameras are those of Tab. 1 . MapA and V-

\Omega

abbreviate MapAnything and VGGT-

\Omega

, and Any3R ∗ is Any3R without its ray map, which assumes an equirectangular grid, and is not ranked.

Figure 5: Zero-shot reconstruction on real fisheye lenses. Per dataset, the two samples from different scenes on which Any3R leads the zero-shot methods shown most (Wid3R † is trained on ScanNet++). Poses and alignment as in Tab. 2 ; the highest $30\%$ of each cloud is removed so the room is seen from above.

Figure 6: Eight central camera models, seven methods, one scene per row. Every method fuses the same three native views with its own predicted poses and is placed in the ground-truth frame by one similarity transform, and the highest $30\%$ of each indoor cloud is removed so the room is seen from above. Column MapAny. is MapAnything and Any3R ∗ is Any3R without its ray map.

Component			M3D		S2D3DS		PanoCity
Ang.	Per- view	Solid- ang.	AUC $\uparrow$	$\overline{\mathrm{R}}_{\text{med}}\downarrow$	AUC $\uparrow$	$\overline{\mathrm{R}}_{\text{med}}\downarrow$	AUC $\uparrow$	$\overline{\mathrm{R}}_{\text{med}}\downarrow$
–	–	✓ ∗	0.619	4.61	0.636	4.18	0.854	0.77
✓	–	–	0.364	9.26	0.381	8.94	0.653	1.88
✓	–	✓	0.380	8.25	0.395	8.50	0.686	1.72
✓	✓	–	0.488	6.36	0.527	6.44	0.633	1.41
✓	✓	✓	0.492	7.13	0.528	6.90	0.669	1.41

Table 4: Module ablation on full sequences under the PanoVGGT protocol (Structured3D, with no multi-view overlap, is not reported). AUC at

15^{\circ}

and

\overline{\mathrm{R}}_{\text{med}}

, the per-sequence median rotation error averaged over sequences; the row marked ∗ has the warm-start architecture and is not directly comparable.

Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

	Matterport3D							Stanford2D3DS
Camera (FoV)	Any3R	$\pi^{3}$ [ 61 ]	MapA [ 25 ]	DA3 [ 34 ]	V- $\Omega$ [ 58 ]	VGGT [ 57 ]	PanoV [ 18 ]	Any3R	$\pi^{3}$	MapA	DA3	V- $\Omega$	VGGT	PanoV
Pinhole (90 ∘ )	0.069	0.109	0.089	0.300	0.161	0.288	0.281	0.083	0.059	0.095	0.325	0.138	0.252	0.284
Brown–C. (90 ∘ )	0.096	0.117	0.150	0.284	0.139	0.320	0.261	0.085	0.120	0.149	0.303	0.184	0.243	0.284
Equidist. (180 ∘ )	0.056	0.166	0.246	0.360	0.215	0.343	0.248	0.061	0.177	0.299	0.363	0.288	0.380	0.270
ERP-Cr. (180 ∘ )	0.051	0.265	0.242	0.267	0.245	0.261	0.331	0.053	0.310	0.296	0.309	0.286	0.287	0.364
MEI (180 ∘ )	0.054	0.163	0.245	0.361	0.212	0.371	0.250	0.064	0.165	0.289	0.365	0.288	0.387	0.281

Appendix

Table 5: Reconstruction accuracy on the Any3R benchmark. Acc

\downarrow

(m) for each camera and dataset, every method with its own predicted poses. In all tables best and second best are shaded, and MapA, V-

\Omega

, PanoV, M3D, S2D3DS and S3D abbreviate MapAnything, VGGT-

\Omega

, PanoVGGT, Matterport3D, Stanford2D3DS and Structured3D.

	Matterport3D							Stanford2D3DS
Camera (FoV)	Any3R	$\pi^{3}$ [ 61 ]	MapA [ 25 ]	DA3 [ 34 ]	V- $\Omega$ [ 58 ]	VGGT [ 57 ]	PanoV [ 18 ]	Any3R	$\pi^{3}$	MapA	DA3	V- $\Omega$	VGGT	PanoV
Pinhole (90 ∘ )	0.052	0.041	0.069	0.261	0.101	0.252	0.266	0.053	0.041	0.062	0.312	0.087	0.239	0.259
Brown–C. (90 ∘ )	0.063	0.072	0.151	0.259	0.126	0.274	0.253	0.069	0.068	0.121	0.297	0.122	0.247	0.287
Equidist. (180 ∘ )	0.051	0.139	0.265	0.328	0.195	0.313	0.230	0.050	0.141	0.296	0.357	0.257	0.383	0.243
ERP-Cr. (180 ∘ )	0.046	0.243	0.212	0.267	0.239	0.243	0.337	0.041	0.268	0.268	0.290	0.245	0.251	0.344
MEI (180 ∘ )	0.048	0.139	0.249	0.302	0.201	0.332	0.236	0.051	0.130	0.268	0.343	0.258	0.376	0.260

Appendix

Table 6: Median accuracy. The median over scenes of the per-scene Acc

\downarrow

(m), the statistic the verdicts of Tab. 5 are decided on.

Method	Matterport3D		Stanford2D3DS		Structured3D		PanoCity
	$90^{\circ}$	$\geq 180^{\circ}$	$90^{\circ}$	$\geq 180^{\circ}$	$90^{\circ}$	$\geq 180^{\circ}$	$90^{\circ}$	$\geq 180^{\circ}$
Any3R (ours)	0.0826	0.0517	0.0838	0.0614	0.7317	0.3449	1.2019	0.8265
$\pi^{3}$ [ 61 ]	0.1132	0.2253	0.0896	0.2439	0.7331	0.5266	1.5954	2.1415
MapAnything [ 25 ]	0.1196	0.2634	0.1221	0.3269	0.7226	0.5527	1.5591	2.5135
DA3 [ 34 ]	0.2922	0.3557	0.3139	0.3738	0.6907	0.5728	5.0245	5.3842
VGGT- $\Omega$ [ 58 ]	0.1501	0.2601	0.1607	0.3243	0.7866	0.6062	1.6416	2.3678

Appendix

Table 7: Accuracy split by field of view. Acc

\downarrow

(m), mean of the cell means of the 2 narrow (

90^{\circ}

) and the 6 wide (

\geq 180^{\circ}

) cameras.

Dataset	Cameras	Any3R	$\pi^{3}$ [ 61 ]	MapA [ 25 ]	DA3 [ 34 ]	V- $\Omega$ [ 58 ]	VGGT [ 57 ]	PanoV [ 18 ]
Matterport3D	Pinhole + KB4 + ERP	0.108	0.343	0.444	0.357	0.377	0.357	0.114
Matterport3D	Brown–C. + Double-Sph. + ERP-Cr.	0.057	0.278	0.254	0.325	0.268	0.355	0.290
Stanford2D3DS	Pinhole + KB4 + ERP	0.116	0.359	0.374	0.399	0.405	0.433	0.133
Stanford2D3DS	Brown–C. + Double-Sph. + ERP-Cr.	0.064	0.244	0.306	0.358	0.341	0.395	0.329
PanoCity	Pinhole + KB4 + ERP	1.282	1.821	1.438	3.242	1.791	2.538	1.085
PanoCity	Brown–C. + Double-Sph. + ERP-Cr.	1.215	2.568	2.329	4.944	2.543	3.842	3.006

Appendix

Table 8: Mixed cameras in one forward pass. Acc

\downarrow

(m) when view

k

of each scene comes from the

k

-th camera listed, each view at its own size on one zero-padded canvas.

		$\pi^{3}$		VGGT		V- $\Omega$		DA3		MapA
Dataset	Cameras	nat.	off.	nat.	off.	nat.	off.	fixed	canvas	fixed	canvas
Matterport3D	Pinhole + KB4 + ERP	0.345	0.328	0.314	0.324	0.354	0.348	0.315	0.351	0.344	0.315
Matterport3D	Brown–C. + Double-Sph. + ERP-Cr.	0.249	0.256	0.313	0.362	0.244	0.251	0.335	0.251	0.233	0.260
Stanford2D3DS	Pinhole + KB4 + ERP	0.378	0.357	0.447	0.424	0.401	0.433	0.378	0.410	0.422	0.379
Stanford2D3DS	Brown–C. + Double-Sph. + ERP-Cr.	0.242	0.243	0.383	0.399	0.300	0.331	0.359	0.346	0.325	0.301
PanoCity	Pinhole + KB4 + ERP	2.513	1.238	4.019	1.844	1.277	1.109	2.831	3.128	1.049	1.801

Appendix

Table 9: Both inputs of the two-input baselines on mixed cameras. Median per-scene Acc

\downarrow

(m); Tab. 8 uses the bold one.

	Matterport3D							Stanford2D3DS
Camera (FoV)	Any3R	$\pi^{3}$ [ 61 ]	MapA [ 25 ]	DA3 [ 34 ]	V- $\Omega$ [ 58 ]	VGGT [ 57 ]	PanoV [ 18 ]	Any3R	$\pi^{3}$	MapA	DA3	V- $\Omega$	VGGT	PanoV
Pinhole (90 ∘ )	3.91	9.84	13.55	56.33	14.94	23.82	35.08	11.00	7.30	23.29	55.28	12.52	23.80	38.99
Brown–C. (90 ∘ )	11.68	14.92	21.13	68.61	14.31	41.34	26.91	11.08	26.68	38.14	64.58	21.61	51.50	33.58
Equidist. (180 ∘ )	8.61	12.20	33.63	55.00	18.92	47.20	13.11	4.82	18.93	43.03	67.73	34.08	49.77	18.56
ERP-Cr. (180 ∘ )	8.40	24.13	36.10	22.10	26.86	30.50	16.66	5.23	33.80	52.67	21.25	35.16	41.16	22.83
MEI (180 ∘ )	1.47	11.62	33.56	59.84	19.45	47.58	11.31	4.92	18.75	43.39	61.79	35.53	52.14	22.03

Appendix

Table 10: Relative pose of the methods in Tab. 5 .

\overline{\mathrm{R}}_{\text{med}}\downarrow

(degrees), each scene’s median rotation error over its view pairs averaged over the scenes, with view pairs of ground-truth translation below

0.01

dropped as in the PanoVGGT code; this table is not ranked.

Figure 7: One Matterport3D scene, all eight central camera models, all seven methods. Identical protocol to Fig. 6 . The scene is fixed, so the vertical axis is field of view alone.

Table 17

	Matterport3D							Stanford2D3DS
FoV	Any3R	DA3 [ 34 ]	VGGT [ 57 ]	$\pi^{3}$ [ 61 ]	V- $\Omega$ [ 58 ]	MapA [ 25 ]	PanoV [ 18 ]	Any3R	DA3	VGGT	$\pi^{3}$	V- $\Omega$	MapA	PanoV
90 ∘	0.083	0.292	0.304	0.113	0.150	0.120	0.271	0.084	0.314	0.247	0.090	0.161	0.122	0.284
150 ∘	0.051	0.330	0.331	0.139	0.176	0.229	0.259	0.053	0.309	0.344	0.133	0.261	0.243	0.281
180 ∘	0.055	0.334	0.329	0.184	0.229	0.250	0.273	0.062	0.355	0.363	0.207	0.297	0.295	0.302
200 ∘	0.054	0.376	0.361	0.177	0.261	0.248	0.242	0.062	0.381	0.378	0.203	0.328	0.317	0.265
360 ∘	0.063	0.366	0.383	0.380	0.364	0.327	0.246	0.081	0.390	0.390	0.393	0.408	0.399	0.275

Appendix

Table 15: Accuracy by field of view on all thirteen cameras. Acc

\downarrow

(m) averaged over the cameras of each field of view.

Figure 8: Projection of the thirteen camera models of Tab. 16 . Left of each pair, a section through the camera centre $O$ shows probe rays spaced $15^{\circ}$ apart in incidence angle $\theta$ , and $30^{\circ}$ apart for ERP, one colour per angle. Each travels from the scene to $O$ (solid) and on to its image point (dashed) on the image plane, a focal length $f$ behind $O$ , so the spacing of the dots is the model’s $r(\theta)$ . ERP and ERP-Crop are seen from above, their image being the sphere of radius $f$ unrolled, and Cylindrical from the side, its image the cylinder wall a distance $f$ in front. Right of each pair, the same probes in 3D, from their image points through $O$ to the unit sphere, with the image’s rows and columns drawn on the sphere in grey. Here $\phi$ and $\lambda$ are azimuth and elevation, $u$ and $v$ image coordinates, $\rho=\tan\theta$ and $d=\sqrt{\sin^{2}\theta+(\xi+\cos\theta)^{2}}$ . Every camera is built as in the benchmark, with the declared field of view reaching half the image height, $k_{1}=-0.18$ and $k_{2}=0.02$ for Brown–Conrady together with the tangential terms $p_{1}=8\times 10^{-4}$ and $p_{2}=-6\times 10^{-4}$ , $\omega=0.9$ for FOV/ATAN, $\xi=1.6$ for MEI, $(k_{1},k_{2},k_{3},k_{4})=(0.04,-0.01,0.002,-0.0003)$ for KB4, $\xi=0.25$ and $\alpha=0.55$ for Double-Sphere, $\alpha=0.55$ and $\beta=1$ for EUCM, and a vertical field of view of $90^{\circ}$ for ERP-Crop and Cylindrical.

	Matterport3D							Stanford2D3DS
Camera (FoV)	Any3R	DA3 [ 34 ]	VGGT [ 57 ]	$\pi^{3}$ [ 61 ]	V- $\Omega$ [ 58 ]	MapA [ 25 ]	PanoV [ 18 ]	Any3R	DA3	VGGT	$\pi^{3}$	V- $\Omega$	MapA	PanoV
Pinhole (90 ∘ )	0.069	0.300	0.288	0.109	0.161	0.089	0.281	0.083	0.325	0.252	0.059	0.138	0.095	0.284
Brown–C. (90 ∘ )	0.096	0.284	0.320	0.117	0.139	0.150	0.261	0.085	0.303	0.243	0.120	0.184	0.149	0.284
FOV/ATAN † (150 ∘ )	0.051	0.330	0.331	0.139	0.176	0.229	0.259	0.053	0.309	0.344	0.133	0.261	0.243	0.281
Equidist. (180 ∘ )	0.056	0.360	0.343	0.166	0.215	0.246	0.248	0.061	0.363	0.380	0.177	0.288	0.299	0.270
Stereogr. † (180 ∘ )	0.060	0.349	0.340	0.144	0.246	0.269	0.262	0.069	0.384	0.399	0.177	0.328	0.295	0.293

Appendix

Table 16: Accuracy on all

52

cells. Acc

\downarrow

(m) per camera, grouped by field of view; the five cameras marked

\dagger

each isolate one axis the original eight leave confounded.

	Matterport3D								Stanford2D3DS
	$\pi^{3}$		VGGT		V- $\Omega$		MapA		$\pi^{3}$		VGGT		V- $\Omega$		MapA
Camera	nat.	off.	nat.	off.	nat.	off.	$-$ K	+K	nat.	off.	nat.	off.	nat.	off.	$-$ K	+K
Pinhole (90 ∘ )	0.050	0.041	0.279	0.252	0.128	0.101	0.125	0.069	0.045	0.041	0.291	0.239	0.118	0.087	0.119	0.062
Brown–C. (90 ∘ )	0.080	0.072	0.295	0.274	0.175	0.126	0.151	–	0.071	0.068	0.338	0.247	0.177	0.122	0.121	–
FOV/ATAN (150 ∘ )	0.119	0.127	0.286	0.279	0.155	0.158	0.221	–	0.115	0.108	0.352	0.333	0.206	0.211	0.236	–
Equidist. (180 ∘ )	0.161	0.139	0.347	0.313	0.211	0.195	0.265	–	0.141	0.141	0.415	0.383	0.257	0.265	0.296	–

Appendix

Table 17: Both inputs of the two-input baselines. Median per-scene Acc

\downarrow

(m) for the native and the official input of

\pi^{3}

, VGGT and VGGT-

\Omega

and for MapAnything without and with pinhole intrinsics; every other table uses the bold one.

Acc $\downarrow$ (m)
	Matterport3D				Stanford2D3DS				Structured3D				PanoCity
Method	KB4	Eqs.	Eqd.	Str.	KB4	Eqs.	Eqd.	Str.	KB4	Eqs.	Eqd.	Str.	KB4	Eqs.	Eqd.	Str.
Any3R	0.050	0.056	0.056	0.060	0.061	0.055	0.061	0.069	0.294	0.310	0.443	0.451	0.863	0.893	0.866	0.917
DA3 [ 34 ]	0.388	0.382	0.360	0.349	0.404	0.364	0.363	0.384	0.563	0.519	0.539	0.613	6.463	6.336	6.656	6.398
VGGT [ 57 ]	0.344	0.371	0.343	0.340	0.368	0.362	0.380	0.399	0.635	0.541	0.623	0.636	6.623	6.280	6.296	5.684
$\pi^{3}$ [ 61 ]	0.175	0.201	0.166	0.144	0.203	0.213	0.177	0.177	0.508	0.467	0.531	0.582	2.394	2.425	2.093	1.991

Appendix

Table 18: Two radial mappings at the same field of view. Acc

\downarrow

(m) of two camera pairs covering the same solid angle (KB4 and Equisolid at

200^{\circ}

, Equidistant and Stereographic at

180^{\circ}

), with the

\overline{\mathrm{R}}_{\text{med}}

(degrees) of the same arms below, so a change of Acc can be read against its change of pose.

Figure 9: Per-point error on the eight camera models. Same cells and same viewpoints as Fig. 6 , with colour giving each point’s distance to the ground truth as a fraction of the scene radius. The colour scale is shared across all cells.

Table 24

Figure 10: One checkpoint with and without the ray map. Ray map is the true ray field and Fallback is the equirectangular one the backbone assumes when none is supplied. Colour is distance to the ground truth as a fraction of the scene radius, on the scale of Fig. 9 . Every camera degrades when the ray map is withdrawn except the equirectangular one in the last row, where the fallback coincides with the true field and nothing changes.

Figure 11: The eight central camera models. The same $360^{\circ}$ panorama imaged through each lens, so every difference in distortion and in coverage comes from the camera model alone. Under each panel: field of view, then the fraction of the sphere that lens observes. Overlaid graticule: equator in red, prime meridian in teal, $45^{\circ}$ meridians and $30^{\circ}$ parallels in white, and the region outside the image circle in light grey.

Figure 12: From one panorama to a native camera frame. Top: the source panorama with the footprints of a $90^{\circ}$ pinhole and a $200^{\circ}$ fisheye marked. Middle: the ray map with direction shown as colour, the validity mask, and the resampled image. Bottom: the same panorama through all eight models. A fresh camera is drawn per view during training.

Figure 13: Acc against field of view, each method normalised by its own

90^{\circ}

value and averaged over the three datasets. The left panel shows all seven general-purpose methods and the right one Any3R and the three that degrade least, on a magnified scale. Below one means the wider lens is the easier one for that method.

Figure 14: Real pinhole imagery on three DL3DV scenes with three views each. Every method fuses its own per-view points with its own predicted poses, one similarity transform fitted on the tracks of the released structure-from-motion reconstruction places each cloud, and each row is rendered from one viewpoint. Sky pixels are found once per input image and removed from every method, and each cloud drops its own farthest $10\%$ of points from its own cameras.

Dataset	Camera (FoV after crop)	Any3R	$\pi^{3}$ [ 61 ]	MapAnything [ 25 ]	DA3 [ 34 ]	VGGT- $\Omega$ [ 58 ]	VGGT [ 57 ]	PanoVGGT [ 18 ]
Matterport3D	Pinhole (53 ∘ )	0.0681	0.0619	0.0749	0.0669	0.0513	0.0927	0.2134
	Brown–Conrady (53 ∘ )	0.0574	0.0488	0.0778	0.0575	0.0509	0.0841	0.2412
	Equidistant (90 ∘ )	0.0622	0.0698	0.1332	0.0865	0.0872	0.1098	0.2178
	KB4 (100 ∘ )	0.0720	0.0793	0.1205	0.0890	0.0976	0.1222	0.2224
	ERP (180 ∘ )	0.0717	0.2182	0.2476	0.2890	0.2660	0.2512	0.2567
Stanford2D3DS	Pinhole (53 ∘ )	0.0521	0.0403	0.0485	0.0391	0.0364	0.0910	0.2362

Appendix

Table 21: Cropping every camera to a narrower field of view. Median Acc

\downarrow

(m) over 10 scenes when each input is centre-cropped to half its angular radius and resized back, scored against the ground-truth points the crop can see.

Figure 15: One camera per row, one crop per column, on Matterport3D. The inset in each cell is the input that produced it, and the pair of numbers is its Acc and the field of view left after cropping. The viewpoint is shared along a row, so a row shows the field of view narrowing and the reconstruction shrinking together.

Figure 16: Accuracy against field of view with the camera held fixed. Each input is centre-cropped to

1.00

0.75

0.50

0.35

and

0.20

of its width and resized back, so only the field of view changes. Reading a curve right to left is reading what a narrower lens costs.

Figure 17: Convergence of an earlier full run. Forty epochs from the original initialisation on eight A100s, evaluated under the paper protocol at

518{\times}1036

on equirectangular input. Dashed lines give the released panoramic baseline on the same protocol; the shaded band is the warm-up.

mean over scenes
	Matterport3D			Stanford2D3DS			Structured3D			PanoCity
FoV	const.	FoV	split	const.	FoV	split	const.	FoV	split	const.	FoV	split
90 ∘	0.066	0.084	0.083	0.084	0.083	0.084	0.783	0.690	0.732	1.185	1.272	1.202
180 ∘	0.042	0.054	0.053	0.049	0.065	0.059	0.334	0.431	0.436	0.734	0.839	0.844
200 ∘	0.044	0.055	0.053	0.054	0.070	0.063	0.233	0.354	0.317	0.805	1.070	0.882
360 ∘	0.047	0.045	0.045	0.069	0.053	0.064	0.126	0.197	0.128	0.645	0.760	0.663

Appendix

Table 22: Accuracy of the three rotary-frequency configurations. Acc

\downarrow

(m) averaged over the cameras of Tab. 5 at each field of view, as the mean and as the median over the scenes of each cell.

3D Reconstruction

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

See all 955 papers

Mar 23, 2026

CAM3R: Camera-Agnostic Model for 3D Reconstruction

Recovering dense 3D geometry from unposed images remains a foundational challenge in computer vision. Current state-of-the-art models are predominantly trained on perspective datasets, which implicitly constrains them to a standard pinhole camera geometry. As a result, these models suffer from significant geometric degradation when applied to wide-angle imagery captured via non-rectilinear optics, such as fisheye or panoramic sensors. To address this, we present CAM3R, a Camera-Agnostic, feed-forward Model for 3D Reconstruction capable of processing images from wide-angle camera models without prior calibration. Our framework consists of a two-view network which is bifurcated into a Ray Module (RM) to estimate per-pixel ray directions and a Cross-view Module (CVM) to infer radial distance with confidence maps, pointmaps, and relative poses. To unify these pairwise predictions into a consistent 3D scene, we introduce a Ray-Aware Global Alignment framework for pose refinement and scale optimization while strictly preserving the predicted local geometry. Extensive experiments on various camera model datasets, including panorama, fisheye and pinhole imagery, demonstrate that CAM3R establishes a new state-of-the-art in pose estimation and reconstruction.

Namitha Guruprasad, Abhay Yadav, Cheng Peng +1

Johns Hopkins University, Baltimore, MD 21218, USA · University of Virginia, Charlottesville, VA 22904, USA

3D Reconstruction Multi-View 3D Reconstruction Feed-Forward 3D Reconstruction Panoramic 3D Reconstruction Camera Pose Estimation

Ask PDF

Oct 22, 2026

OmniRay: Camera-Agnostic Monocular 3D Reconstruction with Composite Ray Fields

Monocular models recover 3D geometry from a single image in one pass, but most assume a pinhole camera and produce distorted geometry on the wide-angle imagery produced by fisheye and panoramic lenses. Camera-aware models remove this assumption by predicting a per-pixel ray field. However, they represent different camera geometries using a single fixed basis to capture both the overall camera projection and its finer variations. We argue that camera geometry can be viewed as having a hierarchy of complexity for representing rays. We present OmniRay, a feed-forward model whose Ray Module is composed of three components: a monotone radial base that captures radially symmetric structure, a spherical-harmonic residual for smooth deviations from it, and a free-form residual for remaining local variation. These components are composed as successive rotations on the sphere, allowing the representation to accommodate diverse camera geometries. Given a single image, OmniRay predicts a ray direction and an up-to-scale radial distance for every pixel, from which a 3D point cloud is recovered without needing camera information at inference time. Extensive experiments across different camera configurations demonstrate that OmniRay establishes a new state-of-the-art in camera-agnostic monocular 3D reconstruction.

Namitha Guruprasad, Abhay Yadav, Cheng Peng +2

Johns Hopkins University · University of Virginia

3D Reconstruction Monocular 3D Reconstruction Feed-Forward 3D Reconstruction 3D Scene Representation

Ask PDF

Oct 12, 2026

Graph3R: Unifying View Graph Initialization with Feed-Forward Reconstruction

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.

Jingwen Yu, Hengli Wang, Jianhao Jiao +3

Hong Kong University of Science and Technology · Southern University of Science and Technology · China Merchants Group, LionRock AI Lab +1

3D Reconstruction Multi-View 3D Reconstruction Feed-Forward 3D Reconstruction Camera Pose Estimation

Ask PDF

Abstract

3D Reconstruction

Explore similar work

CAM3R: Camera-Agnostic Model for 3D Reconstruction

OmniRay: Camera-Agnostic Monocular 3D Reconstruction with Composite Ray Fields

Graph3R: Unifying View Graph Initialization with Feed-Forward Reconstruction