HoloRecon: Incremental Reconstruction of Long-Range Panoramas with Global Consistency

Zhuoxiao Li; Xinyi Liu; Hongye Hou; Jing OU; Kang Du; Shuai Zhang; Dongli Wu; Taoyu Wu

cs.CVOct 27, 2026

HoloRecon: Incremental Reconstruction of Long-Range Panoramas with Global Consistency

Authors: Zhuoxiao Li, Xinyi Liu, Hongye Hou, Jing OU, Kang Du, Shuai Zhang, Dongli Wu, Taoyu Wu,

Organizations: The Hong Kong University of Science and Technology (GuangZhou) · University of Liverpool · Hong Kong Polytechnic University · Xi’an Jiaotong-Liverpool University

Abstract

Long-sequence panoramic reconstruction requires detailed scene geometry and globally consistent camera poses across distant revisits. Panoramic SLAM and incremental SfM rely on feature correspondences, but repeated, similar structures and small partial overlap make reliable long-range connections difficult to establish, leaving accumulated drift unresolved. Feed-forward models offer detailed local geometry, but their high computational costs limit inference window size, necessitating additional constraints for global consistency. This motivates combining continuous feature-based tracking and sparse mapping with rapid, high-quality local reconstruction from feed-forward models to achieve accurate localization and globally consistent dense mapping over long sequences. We therefore propose a panoramic reconstruction framework that couples an ORB-based sparse panoramic mapping frontend with a fine-tuned feed-forward model. Our key insight is that predicted local geometry enables 3D verification of feature matches across distant revisits, yielding long-range camera constraints even between regions. We calibrate predicted depths with sparse landmarks and use verified connections to optimize the global pose graph, updating dense geometry with the corrected camera poses. We also introduce PanoRecon, a seven-scene benchmark with repetitive architecture, large and small loops, and reference camera poses and LiDAR geometry for evaluating global consistency. Our method achieves the lowest mean absolute trajectory error and high F-1 Score among the evaluated methods. Project Page:https://zhuoxiaoli.github.io/HoloRecon_page/

Figures & tables

Figure 1: Overview of the proposed method. Sparse landmarks calibrate Pi3 depths, and registered panoramic cameras place them in the map. These predictions support cross-view ORB correspondences for verifying long-range camera constraints. Global correction and observation fusion refine the sparse map, which recalibrates and repositions the cached dense geometry.

Figure 2: Cross-view evidence for a global connection.

Figure 3: Dataset statistics and GT examples of the PanoRecon benchmark.

Feed-Forward
Method	360LOC			YUTO MMS			PanoRecon
Method	ATE $\downarrow$	RPE-t $\downarrow$	RPE-r $\downarrow$	ATE $\downarrow$	RPE-t $\downarrow$	RPE-r $\downarrow$	ATE $\downarrow$	RPE-t $\downarrow$	RPE-r $\downarrow$
PanoVGGT (ERP)	OOM	OOM	OOM	OOM	OOM	OOM	OOM	OOM	OOM
Scal3R (Pinhole)	5.366	0.124	1.354	187.407	2.911	1.025	9.794	0.282	0.340
SLAM
360DVO (ERP)	0.954	0.026	0.039	68.313	0.183	0.011	2.799	0.104	0.039

Table 1: Quantitative comparison of localization. Mean scene-wise RMSE on 360LOC (4 scenes), YUTO MMS (4 sequences), and PanoRecon (7 scenes). OOM denotes GPU out-of-memory.

Figure 4: Qualitative comparison of localization.

Method	360LOC			YUTO MMS			PanoRecon
Method	P $\uparrow$	R $\uparrow$	F $\uparrow$	P $\uparrow$	R $\uparrow$	F $\uparrow$	P $\uparrow$	R $\uparrow$	F $\uparrow$
Stella-vSLAM-Dense (ERP)	62.42	58.23	58.68	23.78	14.52	17.24	36.61	31.10	31.97
COLMAP (Pinhole)	81.27	44.79	57.25	29.15	13.23	17.28	71.32	51.83	59.75
Ours (ERP)	79.22	52.87	62.04	60.56	40.77	48.71	64.22	62.36	63.16

Table 2: Quantitative comparison of dense reconstruction. Mean scene-wise precision (P), recall (R), and F-score (F) in percent, at 20 cm on 360LOC (4 scenes), YUTO MMS (4 sequences), and PanoRecon (7 scenes). Higher is better.

Figure 5: Point-cloud comparison on the Residence2 and Road scenes of the PanoRecon benchmark.

Backend	Solve (s) $\downarrow$	Total BA (s) $\downarrow$	Speedup $\uparrow$	Reproj. RMSE $\downarrow$
CPU (OpenMP, 32 threads)	19.430	32.649	$1.00\times$	0.8466
GPU (CUDA Schur)	2.244	7.581	$\mathbf{4.31\times}$	0.8420

Table 3: CPU and GPU bundle adjustment.

Table 10

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Visualization of ours PanoRecon benchmark ground-truth LiDAR geometry.

Figure 8: From online submaps to globally consistent geometry. (a) Online pose estimation and local reconstruction. (b) Submap alignment, with distinct colors indicating different submaps. (c) Candidate selection from geometric overlap. (d) Graph correction using verified long-range connections. (e) BA and realignment remove the duplicated building surfaces left by pose drift after online alignment.

Figure 9: Point-cloud comparison of building facades and local structures in the PanoRecon benchmark’s Building1 scene.

Figure 10: Point-cloud comparison in the Residence scene.

Figure 11: Point-cloud comparison in the HPC scene.

Figure 12: Point-cloud comparison in the 360LOC Concourse scene.

Figure 13: Point-cloud comparison in the YUTO MMS Sequence B scene.

Figure 14: Point-cloud comparison in the YUTO MMS Sequence C scene.

Quantity	Requirement
Total inliers, small / large correction	$\geq 60$ / $\geq 100$
Inlier fraction, small / large correction	$\geq 0.35$ / $\geq 0.50$
Unique features per side	$\geq 40$
Supporting keyframes per side	$\geq 3$ , each with $\geq 10$ inliers
Scale / rotation / mean camera displacement	$[2/3,1.5]$ / $\leq 15^{\circ}$ / $\leq 50h$
Bidirectional distance / viewing angle	$\leq 0.3$ / $\leq 1^{\circ}$

Appendix

Table 8: Connection-verification thresholds. Distances use map units, and

h

is the median keyframe spacing.

Ours		COLMAP (Pinhole)
Stage	Time	Stage	Time
Frontend processing	42.71	Feature extraction	7.21
Pi3 inference	38.96	Matching and verification	26.17
Connection matching	44.66	Rig verification	7.39
Connection verification	11.94	Incremental reconstruction + BA	618.52
Pose-graph optimization	10.08	Photometric MVS	227.33

Appendix

Table 9: Key-stage runtime on YUTO MMS Sequence C. Times are in minutes, excluding separate preprocessing and export.

Configuration	ATE $\downarrow$	RPE-t $\downarrow$	RPE-r $\downarrow$	Precision $\uparrow$	Recall $\uparrow$	F-score $\uparrow$
	(m)	(m)	( ∘ )	(%)	(%)	(%)
Default SL(4)	Failed at submap 57/59
Adapted SE(3)	26.831	0.430	3.616	2.31	45.83	4.40
Without loop factors SE(3)	29.970	0.294	0.112	4.55	51.09	8.35
Ours submap alignment Sim(3)	4.203	0.107	0.149	10.60	78.14	18.67

Appendix

Table 10: Adapt PanoVGGT with VGGT-SLAM on Residence2.

Figure 15: Applying Submap Alignment to PanoVGGT

Figure 16: Reconstruction Result on 360loc Dataset Hall Scene.

Figure 17: Reconstruction Result on Ours PanoRecon Benchmark Residence1 Scene.

Figure 18: Reconstruction Result on Ours PanoRecon Benchmark Road Scene.

Figure 19: Reconstruction Result on Ours PanoRecon Benchmark Residence2 Scene.

Figure 20: Reconstruction Result on YUTO MMS Benchmark Sequence B Scene.

Figure 21: Reconstruction Result on YUTO MMS Benchmark Sequence C Scene.

Atrium					Concourse
Method	$n$	ATE $\downarrow$	RPE-t $\downarrow$	RPE-r $\downarrow$	Method	$n$	ATE $\downarrow$	RPE-t $\downarrow$	RPE-r $\downarrow$
Scal3R	581	3.322	0.095	0.841	Scal3R	491	5.717	0.130	0.611
360DVO	581	0.654	0.044	0.034	360DVO	491	0.779	0.011	0.044
Open-vSLAM	580	0.255	0.691	22.408	Open-vSLAM	471	0.081	0.670	20.111
Stella-vSLAM	285	1.209	0.083	0.059	Stella-vSLAM	262	0.172	0.037	0.298
COLMAP (ERP)	581	0.201	0.043	0.033	COLMAP (ERP)	491	0.131	0.007	0.038

Appendix

Table 11: 360LOC localization by scene. ATE and RPE-t are in meters; RPE-r is in degrees.

n

is the number of associated poses (scene panels) or available/expected scenes (Mean). All errors are RMSE; lower is better.

Sequence A					Sequence B
Method	$n$	ATE $\downarrow$	RPE-t $\downarrow$	RPE-r $\downarrow$	Method	$n$	ATE $\downarrow$	RPE-t $\downarrow$	RPE-r $\downarrow$
Scal3R	234	24.591	0.301	0.409	Scal3R	2794	307.710	0.948	0.381
360DVO	234	0.244	0.006	0.010	360DVO	2794	135.769	0.277	0.013
Open-vSLAM	235	35.215	15.562	56.725	Open-vSLAM	2794	479.963	37.245	1.025
Stella-vSLAM	235	0.370	0.026	0.060	Stella-vSLAM	2794	252.571	1.074	0.223
COLMAP (ERP)	234	0.161	0.008	0.014	COLMAP (ERP)	2794	412.900	2.060	0.197

Appendix

Table 12: YUTO MMS localization by scene. ATE and RPE-t are in meters; RPE-r is in degrees.

n

is the number of KF poses or available/expected scenes (Mean).

Building 1					Building 2
Method	$n$	ATE $\downarrow$	RPE-t $\downarrow$	RPE-r $\downarrow$	Method	$n$	ATE $\downarrow$	RPE-t $\downarrow$	RPE-r $\downarrow$
Scal3R	765	14.643	0.281	0.182	Scal3R	419	3.537	0.223	0.128
360DVO	765	2.726	0.240	0.023	360DVO	419	4.548	0.175	0.047
Open-vSLAM	765	3.387	2.836	34.833	Open-vSLAM	418	5.764	3.636	0.135
Stella-vSLAM	765	2.506	0.411	0.646	Stella-vSLAM	418	5.018	1.024	0.121
COLMAP (ERP)	765	10.831	15.002	8.499	COLMAP (ERP)	419	5.369	0.202	0.048

Appendix

Table 13: PanoRecon localization by scene. ATE and RPE-t are in meters; RPE-r is in degrees.

n

is the number of associated poses (scene panels) or available/expected scenes (Mean). All errors are RMSE; lower is better.

Scene	Stella-vSLAM-Dense			COLMAP (Pinhole)			Ours
Scene	P $\uparrow$	R $\uparrow$	F $\uparrow$	P $\uparrow$	R $\uparrow$	F $\uparrow$	P $\uparrow$	R $\uparrow$	F $\uparrow$
Atrium	47.83	76.16	58.76	67.84	46.63	55.27	65.51	66.92	66.20
Concourse	77.73	59.92	67.67	85.15	54.85	66.72	83.88	52.27	64.40
Hall	71.55	49.30	58.38	77.85	34.27	47.59	85.95	36.60	51.34
Piatrium	52.58	47.53	49.93	94.22	43.40	59.43	81.57	55.71	66.20

Appendix

Table 14: 360LOC dense reconstruction by scene. P/R/F in percent at 20 cm. Colors rank methods within each scene and metric.

Scene	Stella-vSLAM-Dense			COLMAP (Pinhole)			Ours
Scene	P $\uparrow$	R $\uparrow$	F $\uparrow$	P $\uparrow$	R $\uparrow$	F $\uparrow$	P $\uparrow$	R $\uparrow$	F $\uparrow$
Sequence A	46.13	39.25	42.41	88.09	49.94	63.74	88.77	62.81	73.56
Sequence B	18.74	4.19	6.85	10.13	0.81	1.50	48.63	32.51	38.97
Sequence C	13.42	7.10	9.29	9.85	1.36	2.39	66.52	43.20	52.38
Sequence D	16.84	7.51	10.39	8.51	0.81	1.48	38.31	24.56	29.93

Appendix

Table 15: YUTO MMS dense reconstruction by scene. P/R/F in percent at 20 cm. Colors rank methods within each scene and metric.

Scene	Stella-vSLAM-Dense			COLMAP (Pinhole)			Ours
Scene	P $\uparrow$	R $\uparrow$	F $\uparrow$	P $\uparrow$	R $\uparrow$	F $\uparrow$	P $\uparrow$	R $\uparrow$	F $\uparrow$
Building 1	22.72	37.68	28.34	86.74	71.51	78.39	70.69	68.87	69.77
Building 2	49.98	55.64	52.66	67.20	62.37	64.69	64.06	58.56	61.19
Gym	34.40	10.87	16.52	58.60	41.62	48.67	66.03	53.86	59.33
HPC	51.50	30.84	38.58	84.42	55.84	67.21	65.74	66.36	66.05
Residence	50.17	52.52	51.32	89.47	56.71	69.42	76.00	74.44	75.22

Appendix

Table 16: PanoRecon dense reconstruction by scene. P/R/F in percent at 20 cm. Colors rank methods within each scene and metric.

Panoramic 3D Reconstruction

Panoramic 3D reconstruction recovers navigable, geometrically consistent scenes from 360-degree images, video, and sparse multi-view captures.

See all 32 papers

Holo360D: A Large-Scale Real-World Dataset with Continuous Trajectories for Advancing Panoramic 3D Reconstruction and Beyond
Apr 24, 2026Jing Ou, Zidong Cao, Yinrui Ren +6Panoramic 3D Reconstruction 3D Reconstruction
1The Hong Kong University of Science and Technology (Guangzhou) · 2South China Normal University
PanoImager: Geometry-Guided Novel View Synthesis and Reconstruction from Sparse Panoramic Views
Jun 25, 2026Zhisong Xu, Takeshi OishiMulti-View Consistency 3D Gaussian Splatting
Institute of Industrial Science, The University of Tokyo, Tokyo, Japan
Glob3R: Global Structure-from-Motion with 3D Foundation Models
Jul 10, 2026Junyuan Deng, Heng Li, Kejie Qiu +7Bundle Adjustment Multi-View 3D Reconstruction
The Hong Kong University of Science and Technology · Tongyi Lab, Alibaba Group · Nanjing University +1

Apr 24, 2026J/K move · Enter open · S save

Holo360D: A Large-Scale Real-World Dataset with Continuous Trajectories for Advancing Panoramic 3D Reconstruction and Beyond

Jing Ou, Zidong Cao, Yinrui Ren, Zhuoxiao Li+5

1The Hong Kong University of Science and Technology (Guangzhou) · 2South China Normal University

Panoramic 3D Reconstruction 3D Reconstruction Multi-View 3D Reconstruction Feed-Forward 3D Reconstruction

Open paper AskPDF arXiv

While feed-forward 3D reconstruction models have advanced rapidly, they still exhibit degraded performance on panoramas due to spherical distortions. Moreover, existing panoramic 3D datasets are predominantly collected with 360 cameras fixed at discrete locations, resulting in discontinuous trajectories. These limitations critically hinder the development of panoramic feed-forward 3D reconstruction, especially for the multi-view setting. In this paper, we present Holo360D, a comprehensive dataset containing 109,495 panoramas paired with registered point clouds, meshes, and aligned camera poses. To our knowledge, Holo360D is the first large-scale dataset that provides continuous panoramic sequences with accurately aligned high-completeness depth maps. The raw data are initially collected using a 3D laser scanner coupled with a 360 camera. Subsequently, the raw data are processed with both online and offline SLAM systems. Furthermore, to enhance the 3D data quality, a post-processing pipeline tailored for the 360 dataset is proposed, including geometry denoising, mesh hole filling, and region-specific remeshing. Finally, we establish a new benchmark by fine-tuning 3D reconstruction models on Holo360D, providing key insights into effective fine-tuning strategies. Our results demonstrate that Holo360D delivers superior training signals and provides a comprehensive benchmark for advancing panoramic 3D reconstruction models. Datasets and Code will be made publicly available.

Abstract

Panoramic 3D Reconstruction

Explore similar work

Holo360D: A Large-Scale Real-World Dataset with Continuous Trajectories for Advancing Panoramic 3D Reconstruction and Beyond

PanoImager: Geometry-Guided Novel View Synthesis and Reconstruction from Sparse Panoramic Views

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