FullCircle: Effortless 3D Reconstruction from Casual 360$^\circ$ Captures

Yalda Foroutan; Ipek Oztas; Daniel Rebain; Aysegul Dundar; Kwang Moo Yi; Lily Goli; Andrea Tagliasacchi

cs.CVMar 23, 2026

FullCircle: Effortless 3D Reconstruction from Casual 360 $^\circ$ Captures

Authors: Yalda Foroutan, Ipek Oztas, Daniel Rebain, Aysegul Dundar, Kwang Moo Yi, Lily Goli, Andrea Tagliasacchi

Organizations: Simon Fraser University · Bilkent University · University of British Columbia · University of Toronto

Abstract

Radiance fields have emerged as powerful tools for 3D scene reconstruction. However, casual capture remains challenging due to the narrow field of view of perspective cameras, which limits viewpoint coverage and feature correspondences necessary for reliable camera calibration and reconstruction. While commercially available 360 $^\circ$ cameras offer significantly broader coverage than perspective cameras for the same capture effort, existing 360 $^\circ$ reconstruction methods require special capture protocols and pre-processing steps that undermine the promise of radiance fields: effortless workflows to capture and reconstruct 3D scenes. We propose a practical pipeline for reconstructing 3D scenes directly from raw 360 $^\circ$ camera captures. We require no special capture protocols or pre-processing, and exhibit robustness to a prevalent source of reconstruction errors: the human operator that is visible in all 360 $^\circ$ imagery. To facilitate evaluation, we introduce a multi-tiered dataset of scenes captured as raw dual-fisheye images, establishing a benchmark for robust casual 360 $^\circ$ reconstruction. Our method significantly outperforms not only vanilla 3DGS for 360 $^\circ$ cameras but also robust perspective baselines when perspective cameras are simulated from the same capture, demonstrating the advantages of 360 $^\circ$ capture for casual reconstruction. Additional results are available at: https://theialab.github.io/fullcircle

Figures & tables

Figure 1: Teaser – With 360 ∘ cameras we can recover 3D scenes far more effectively than with traditional perspective cameras, resulting in substantial gains in novel view synthesis (left: perspective vs. 360 ∘ ). However, casual 360 ∘ captures unavoidably include the camera operator, which degrades reconstruction quality if left unaddressed (e.g. 3DGRT [ 35 ] ). Our FullCircle method overcomes this challenge, enabling fast, high-quality scene capture using 360 ∘ cameras.

Scene	#training	#test	#total	#w/ mask	#w/o mask	duration	difficulty
Room 1	442 / 117	38	597	–	8	88s / 23s	Easy
Flat 1	440 / 153	46	639	–	56	88s / 31s	Easy
Flat 2	434 / 107	26	567	–	1	87s / 21s	Easy
Room 2	252 / 82	28	362	–	–	50s / 16s	Easy
Room 3	257 / 100	30	387	1	104	51s / 20s	Medium
Lab	431 / 116	34	581	–	3	86s / 23s	Medium

Table 1: We collect a dataset of dual-fisheye scene captures with different levels of difficulty. We report the number of frames where the capturer is static/dynamic, and the number of COLMAP pose failures (w/ and w/o transient masks). The full image set is released to support future work on robust calibration.

Figure 4: Capturer mask estimation – Our masking pipeline first detects the capturer using YOLOv8 [ 20 ] and SAMv2 [ 37 ] ( red masks). These initial masks roughly localize the capturer but include missing or incorrect regions that can degrade reconstruction. We then re-center the omni images on the detected capturer, render synthetic fisheyes, and re-run SAMv2 with a center-point prompt to obtain refined, temporally consistent masks ( purple masks), which are mapped back to the original dual-fisheye inputs.

Figure 5: Dual-fisheye vs. perspective capture – Using dual-fisheye images for both calibration and reconstruction ( – ) yields higher-quality novel views and stable geometry than when reconstruction ( – ), or both calibration and reconstruction ( – ), relies on perspective images. This advantage becomes most apparent when moving beyond the in-distribution training trajectory ( purple cameras) toward out-of-distribution test views ( yellow camera), where ( – ) and ( – ) exhibit a gradual degradation in reconstruction quality.

Figure 6: Effect of undistortion – Using fisheye frames with FullCircle yields higher-quality reconstructions, both qualitatively and quantitatively, compared to perspective baselines—robust and non-robust—that operate on preprocessed, undistorted images.

Figure 7: Fisheye reconstruction – Our method reliably removes human distractors and reconstructs high-quality 3D scenes. The table reports PSNR, where our method outperforms both robust and non-robust radiance-field baselines. OLRF renderings are mapped to fisheye frames for qualitative comparison only.

Figure 8: Camera operator segmentation – We compare our masking approach for removing the camera operator with OmniSAM, a recent segmentation method for omnidirectional images, both qualitatively and quantitatively.

Figure 9: Synthetic fisheye – Without re-centering the fisheye images synthetically, SAMv2 [ 37 ] misses the capturer in some frames, resulting in a noisy reconstruction.

Figure 10: Limitations – We find that 360 ∘ coverage provides enough data redundancy to allow our pipeline to robustly handle other moving distractors (right), but our pipeline struggles when other distractors stay still for long stretches (left).

Method	Robust	Room 1	Flat 1	Flat 2	Room 2	Room 3	Lab	Lounge	Persons	Dark	Mean
3DGS [ 21 ]	✗	0.90	0.83	0.85	0.90	0.85	0.80	0.72	0.82	0.87	0.84
SLS-MLP [ 41 ]	✓	0.90	0.83	0.86	0.91	0.86	0.80	0.71	0.83	0.89	0.84
NOTG [ 40 ]	✓	0.80	0.69	0.86	0.80	0.80	0.70	0.57	0.84	0.84	0.77
FullCircle (Ours)	✓	0.92	0.86	0.89	0.93	0.88	0.83	0.75	0.86	0.90	0.87

Table 2: Perspective – SSIM

\uparrow

metric for reconstruction quality using undistorted fisheye images vs. ours.

Method	Robust	Room 1	Flat 1	Flat 2	Room 2	Room 3	Lab	Lounge	Persons	Dark	Mean
3DGS [ 21 ]	✗	0.22	0.35	0.26	0.25	0.31	0.30	0.26	0.37	0.30	0.29
SLS-MLP [ 41 ]	✓	0.19	0.31	0.22	0.23	0.28	0.30	0.28	0.33	0.25	0.27
NOTG [ 40 ]	✓	0.31	0.50	0.23	0.41	0.34	0.41	0.45	0.30	0.28	0.36
FullCircle (Ours)	✓	0.19	0.30	0.18	0.21	0.26	0.27	0.23	0.31	0.25	0.25

Table 3: Perspective – LPIPS

\downarrow

metric for reconstruction quality using undistorted fisheye images vs. ours.

Method	Robust	Room 1	Flat 1	Flat 2	Room 2	Room 3	Lab	Lounge	Persons	Dark	Mean
3DGRT [ 35 ]	✗	0.94	0.90	0.91	0.93	0.91	0.89	0.85	0.89	0.91	0.90
SLS-MLP [ 41 ]	✓	0.94	0.91	0.93	0.94	0.92	0.90	0.85	0.90	0.92	0.91
OLRF [ 6 ]	✓	0.82	0.74	0.81	0.82	0.78	0.68	0.63	0.75	0.81	0.76
NOTG [ 40 ]	✓	0.86	0.73	0.83	0.92	0.86	0.76	0.69	0.84	0.81	0.81
FullCircle (Ours)	✓	0.95	0.91	0.94	0.95	0.92	0.90	0.87	0.91	0.92	0.92

Table 4: Fisheye – SSIM

\uparrow

metric for reconstruction quality using fisheye images. For OmniLocalRF, we report results using omnidirectional images.

Method	Robust	Room 1	Flat 1	Flat 2	Room 2	Room 3	Lab	Lounge	Persons	Dark	Mean
3DGRT [ 35 ]	✗	0.10	0.19	0.12	0.14	0.16	0.16	0.12	0.20	0.16	0.15
SLS-MLP [ 41 ]	✓	0.09	0.17	0.10	0.11	0.15	0.14	0.12	0.19	0.14	0.14
OLRF [ 6 ]	✓	0.40	0.52	0.42	0.41	0.49	0.50	0.48	0.47	0.46	0.46
NOTG [ 40 ]	✓	0.19	0.47	0.22	0.14	0.23	0.30	0.35	0.25	0.30	0.27
FullCircle (Ours)	✓	0.08	0.16	0.09	0.11	0.15	0.14	0.10	0.18	0.14	0.13

Table 5: Fisheye – LPIPS

\downarrow

metric for reconstruction quality using fisheye images. For OmniLocalRF, we report results using omnidirectional images.

Figure 11: Camera operator segmentation – Additional qualitative comparison between OmniSAM and our method for camera operator segmentation.

Method	Room 1	Flat 1	Flat 2	Room 2	Room 3	Lab	Lounge	Persons	Dark	Mean
DINOv3 [ 47 ]	32.59	29.21	30.50	31.61	29.87	29.43	24.35	30.86	27.99	29.60
FullCircle (Ours)	32.64	29.11	30.55	31.63	29.91	29.54	24.40	30.99	27.79	29.62

Table 6: DINOv3 – Reconstruction quality reported in PSNR

\uparrow

metric for DINOv3 masking alternative to SAMv2.

Method	COLMAP	Room 1	Flat 1	Flat 2	Room 2	Room 3	Lab	Lounge	Persons	Dark	Mean
3DGRT [ 35 ]	Unmasked	27.56	25.75	24.99	27.88	27.48	26.90	23.43	failed	25.51	-
FullCircle (Ours)	Unmasked	32.20	28.31	30.20	31.30	28.80	28.68	24.38	failed	27.77	-
3DGRT [ 35 ]	Masked	27.37	26.81	25.09	27.92	28.41	26.66	23.41	27.56	25.71	26.55
FullCircle (Ours)	Masked	32.64	29.11	30.55	31.63	29.91	29.54	24.40	30.99	27.79	29.62

Table 7: Unmasked calibration – Reconstruction performance reported in PSNR

\uparrow

metric using unmasked calibration.

Figure 12: Ricoh360 dataset – Quantitative comparison on the Ricoh360 dataset [ 5 ] together with qualitative results of our method. Our approach achieves higher PSNR $\uparrow$ in all scenes and reliably removes the static camera holder. Regions underneath the holder remain empty in the reconstruction, since they are unobserved in the input, rather than producing visible artifacts.

3D Reconstruction Benchmarks

See all 15 papers

Mar 23, 2026

4DGS360: 360° Gaussian Reconstruction of Dynamic Objects from a Single Video

We introduce 4DGS360, a diffusion-free framework for 360

^{\circ}

dynamic object reconstruction from casual monocular video. Existing methods often fail to reconstruct consistent 360

^{\circ}

geometry, as their heavy reliance on 2D-native priors causes initial points to overfit to visible surface in each training view. 4DGS360 addresses this challenge through a advanced 3D-native initialization that mitigates the geometric ambiguity of occluded regions. Our proposed 3D tracker, AnchorTAP3D, produces reinforced 3D point trajectories by leveraging confident 2D track points as anchors, suppressing drift and providing reliable initialization that preserves geometry in occluded regions. This initialization, combined with optimization, yields coherent 360

^{\circ}

4D reconstructions. We further present iPhone360, a new benchmark where test cameras are placed up to 135

^{\circ}

apart from training views, enabling 360

^{\circ}

evaluation that existing datasets cannot provide. Experiments show that 4DGS360 achieves state-of-the-art performance on the iPhone360, iPhone, and DAVIS datasets, both qualitatively and quantitatively.

Jae Won Jang, Yeonjin Chang, Wonsik Shin +2

Seoul National University, Seoul, South Korea

Dynamic Scene Reconstruction Panoramic 3D Reconstruction 3D Gaussian Splatting 3D Point Tracking

Ask PDF

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

Aug 20, 2026

Sparse Light Field Sampling Improves Casual 3D and 4D Reconstruction

Many consumer smartphones, stereo cameras, and light field cameras record multiple synchronized viewpoints in a single exposure event. However, novel view synthesis pipelines commonly use only a monocular stream and rely on camera motion or learned priors to obtain angular coverage. In this paper, we ask: why do we use only one viewpoint? We analyze sensor-limited multi-view, where one sensor trades off spatial and angular resolution, and exposure-limited multi-view, where multiple sensors on one commodity device observe each event simultaneously. We introduce a new dataset incorporating three types of commodity multi-view cameras, and evaluate sparse-view 3DGS and 4DGS baselines measuring reconstruction quality as a function of number of exposures and angle between extreme views. Our results demonstrate that using multiple cameras, even with a low baseline, significantly improves reconstruction quality in single-shot, few-shot, and casual video settings. In addition, under a fixed sensor budget, angular sampling improves reconstruction when exposures are scarce despite lower spatial resolution. The gains are most pronounced for single-shot and dynamic scenes, where a stationary monocular camera lacks the angular diversity to recover scene geometry and motion.

Shamus Li, Ruiming Cao, Laura Waller +2

Cornell University, USA · Adobe, USA · University of California, Berkeley, USA +1

3D Reconstruction Sparse-View 4D Reconstruction Sparse-View 3D Reconstruction 4D Reconstruction Dynamic Scene Reconstruction 3D Gaussian Splatting

Ask PDF

Abstract

3D Reconstruction Benchmarks

Explore similar work

4DGS360: 360° Gaussian Reconstruction of Dynamic Objects from a Single Video

CAM3R: Camera-Agnostic Model for 3D Reconstruction

Sparse Light Field Sampling Improves Casual 3D and 4D Reconstruction