A coupled sparse–dense framework
Incremental panoramic mapping and feed-forward geometry share camera poses and landmark-based depth calibration.
PANORAMIC MAPPING & 3D RECONSTRUCTION
HoloRecon Incremental Reconstruction of Long-Range
Panoramas with Global Consistency
1 The Hong Kong University of Science and Technology (Guangzhou)
2 University of Liverpool
3 Hong Kong Polytechnic University
4 Xi'an Jiaotong-Liverpool University
Contact: zhuoxiaoli@hkust-gz.edu.cn
360° panoramic input15 evaluated scenes / sequences7 scenes in the new PanoRecon benchmark
EXPLORE / INTERACTIVE 3DPOINT-CLOUD RECONSTRUCTIONS
Interactive Point Clouds
For browser visualization, Dense point clouds are downsampled to approximately 10 million points per scene.
Inspect scene geometry in 3D. Rotate, zoom, and adjust the rendered point size.
Display settings
Dense and Sparse both display all available points by default. Choose a lower budget for a lighter view; this reloads the current cloud.
Residence 2 loads automatically in Dense mode, with all source points displayed.
01 / OVERVIEWPROBLEM & KEY INSIGHT
Local geometry as evidence
for global consistency.
Sparse mapping anchors predicted geometry.
Predicted geometry verifies distant revisits.
Long-range panoramic reconstruction requires detailed geometry and globally consistent camera poses. Repetitive architecture and limited overlap make distant revisits difficult to verify, while feed-forward models provide local geometry within bounded inference windows.
HoloRecon couples an ORB-based sparse panoramic mapping frontend with a fine-tuned Pi3 model. Sparse landmarks calibrate predicted depths; the resulting geometry supports 3D verification of feature correspondences across distant revisits. Verified connections drive pose-graph optimization, landmark fusion, and GPU-accelerated panoramic bundle adjustment. Cached dense predictions are then updated using corrected camera poses and recalibrated depths, without repeating model inference.
Geometry-verified long-range connections
Cross-view 3D evidence establishes camera constraints without requiring shared sparse landmarks between revisited regions.
The PanoRecon benchmark
Seven real-world scenes with repeated architecture, distant revisits, reference camera poses, and LiDAR geometry.
02 / METHODLOCAL PREDICTIONS & GLOBAL REFINEMENT
Method Overview
Sparse and dense estimates are refined through a shared geometric representation.
Calibrate local geometry
An ORB-based ERP frontend estimates poses and sparse landmarks. Pi3, fine-tuned on Holo360D, predicts local geometry from perspective views. Landmark-supported window scales and per-frame corrections anchor these depths to the sparse map.
Establish long-range constraints
Spatial overlap and visibility propose candidate submaps. Depth-supported ORB correspondences estimate a common Sim(3) transform, verified across multiple keyframes using bidirectional distance and bearing consistency.
Update sparse and dense maps
Pose-graph correction and landmark fusion establish cross-region observations for GPU-accelerated panoramic bundle adjustment. Corrected source-camera poses and refitted depth scales update cached points without rerunning Pi3.
GEOMETRY-SUPPORTED LOOP VERIFICATIONVerify a shared transform
Verify a shared transform
across multiple views.
Similar facades can produce plausible image matches at different locations. HoloRecon checks whether one geometric transform explains correspondences across multiple keyframes on both sides of a revisit.
Large corrections receive additional validation using separate frame groups and neighboring historical submaps before their camera constraints enter global optimization.
Implementation
16 panoramic keyframes per submap, with 4 shared keyframes between consecutive submaps. Experiments use a single NVIDIA RTX PRO 6000 Blackwell Workstation Edition GPU.
03 / BENCHMARKINTRODUCING PANORECON
PanoRecon
Evaluating global consistency under repetitive architecture and long-range revisits.
Panoramic videos and RTK-assisted LiDAR are captured independently along the same planned routes. Large and small loops, repeated architectural patterns, and limited overlap at distant revisits test the ability to correct accumulated drift.
- Scenes
- 7
- Mean video duration
- 11.49 min
- Mean trajectory length
- 0.87 km (0.26–1.69 km)
- Reference frame sampling
- 0.1 s
- Reference reprojection RMSE
- 0.1917 px
Reference camera poses and dense LiDAR geometry share a metric coordinate system. The benchmark supports trajectory and dense reconstruction evaluation.
04 / EXPERIMENTS360LOC · YUTO MMS · PANORECON
Quantitative Evaluation
Lowest mean ATE and highest mean dense F-score among the evaluated methods on all three benchmarks.
PANORECON · MEAN ATE ↓1.697 mAcross 7 scenes
YUTO MMS · MEAN F-SCORE ↑48.71%Dense reconstruction at 20 cm
PANORECON · MEAN F-SCORE ↑63.16%Dense reconstruction at 20 cm
| Method | 360LOC 4 scenes | YUTO MMS 4 sequences | PanoRecon 7 scenes |
|---|---|---|---|
| PanoVGGT (ERP) | OOM | OOM | OOM |
| Scal3R (Pinhole) | 5.366 | 187.407 | 9.794 |
| 360DVO (ERP) | 0.954 | 68.313 | 2.799 |
| Open-vSLAM (ERP) | 0.441 | 281.793 | 7.399 |
| Stella-vSLAM (ERP) | 0.706 | 82.173 | 7.424 |
| COLMAP (ERP) | 0.355 | 124.490 | 3.510 ‡ |
| COLMAP (Pinhole) | 0.399 | 113.317 | 1.736 |
| Spherical-SfM (ERP) | 0.491 | 21.186 | 11.779 |
| HoloRecon (Ours, ERP) | 0.270 | 3.527 | 1.697 |
Camera poses are evaluated after one global Sim(3) alignment. ‡ COLMAP (ERP) reports results on 6 of 7 PanoRecon scenes. OOM: GPU out-of-memory under the manuscript's single-GPU input settings.
| Method | 360LOC 4 scenes | YUTO MMS 4 sequences | PanoRecon 7 scenes |
|---|---|---|---|
| Stella-vSLAM-Dense (ERP) | 58.68 | 17.24 | 31.97 |
| COLMAP (Pinhole) | 57.25 | 17.28 | 59.75 |
| HoloRecon (Ours, ERP) | 62.04 | 48.71 | 63.16 |
This table summarizes F-score from the manuscript's dense reconstruction comparison. Precision, recall, and per-scene breakdowns are reported in the manuscript.
Reported results from the manuscript. Dataset means are computed scene-wise.
Qualitative Comparisons
Trajectory estimates, reconstructed surfaces, and the effect of global refinement.
Reconstructed geometry on PanoRecon Residence 2 and Road. Enlarged regions show local surface structure relative to the LiDAR reference.
Estimated camera trajectories and reference trajectories. Long-range connections constrain accumulated drift across revisits.
Residence 2 ablation: trajectory and reconstruction changes with global connections and joint refinement.
Effect of Global Constraints
On the Residence 2 full-trajectory ablation, pose-graph correction, landmark fusion, and bundle adjustment progressively reduce ATE. The full configuration reduces ATE from 1.358 m to 0.708 m.
| Configuration | ATE (m) ↓ | RPE-t (m) ↓ |
|---|---|---|
| No connection | 1.358 | 0.061 |
| PG | 0.977 | 0.060 |
| PG + BA | 0.922 | 0.057 |
| PG + fusion | 0.877 | 0.060 |
| PG + fusion + BA | 0.708 | 0.048 |
05 / VIDEO DEMONSTRATIONS9 VIDEOS · METHOD & DATASET COLLECTIONS
Reconstruction in Detail
Method demonstrations and scene visualizations, organized by benchmark.
Unable to play this video. Open the MP4 file.
METHOD · PANORECON
02:30Online Reconstruction and Depth Alignment
Local submap construction and depth alignment on PanoRecon Residence 2, followed by the final reconstruction.
06 / SCENE GALLERYGLOBAL RECONSTRUCTIONS & LOCAL DETAILS
Reconstruction Gallery
Selected results from PanoRecon, YUTO MMS, and 360LOC. Select a figure to inspect the original.