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, Siyuan Feng, Xiaohui Zhu, Wufan ZhaoarXiv 2026 · arXiv

Reconstructing a 3D scene from a long sequence of panoramic images is a bit like trying to solve a massive, distorted jigsaw puzzle while walking. You have the full 360-degree view, but as you travel kilometers away and eventually return to a previous spot, "drift"—the accumulation of small errors in estimated position—makes it nearly impossible to align your new observations with the old ones.

The researchers behind HoloRecon have introduced a new framework that tackles this problem by cleverly combining the reliability of traditional feature-based SLAM with the rich geometric priors of modern "feed-forward" deep learning models.

TL;DR

HoloRecon is a panoramic 3D reconstruction framework that achieves global consistency over long-range sequences (up to 1.2 km) by coupling sparse feature tracking with dense feed-forward geometry. Its core innovation is using predicted local 3D geometry to verify feature matches across distant revisits, effectively "locking" the global map and resolving trajectory drift.

What Problem Does This Paper Address?

Large-scale panoramic reconstruction faces a "best of both worlds" dilemma.

  1. Traditional SLAM/SfM: Systems like ORB-SLAM are great at tracking motion incrementally, but they struggle with "loop closure" in complex environments. If two buildings look identical (repetitive architecture) or if the overlap between views is small, the system might fail to realize it has returned to a known location, or worse, create a "false loop" that ruins the map.
  2. Feed-Forward Models: Recent deep learning models can predict incredibly detailed local geometry from just a few images. However, they are computationally expensive and usually operate in small "windows." Without a global backbone, these local submaps don't align correctly over long distances, leading to "ghosting" or duplicated surfaces.

HoloRecon addresses the challenge of maintaining global consistency—ensuring that when you revisit a location after a 1km loop, the reconstructed walls and floors align perfectly rather than appearing as offset layers.

Methods

The HoloRecon pipeline is an elegant "sandwich" of sparse and dense techniques:

  • Frontend Tracking: It starts with an ORB-based sparse mapping system that provides an initial estimate of the camera trajectory and a set of sparse 3D landmarks.
  • Local Dense Reconstruction: It uses a fine-tuned version of the $\pi^3$ model to predict dense depth maps for small windows of keyframes. These local "submaps" are anchored to the sparse SLAM landmarks to establish a common scale.
  • Geometric Loop Verification: This is the key innovation. When the system detects a potential revisit, it doesn't just look at 2D image pixels. It uses the predicted 3D geometry to perform 3D verification. By back-projecting features into 3D space using the predicted depths, the system can calculate a precise similarity transform (rotation, translation, and scale) between distant regions.
  • Global Refinement: These verified 3D connections are fed into a global pose-graph optimization and a custom, GPU-accelerated panoramic Bundle Adjustment (BA).
  • Efficient Updating: Once the global trajectory is corrected, the system repositions the cached dense points. Crucially, it does this without re-running the expensive deep-learning inference, making the global update much faster.
The HoloRecon pipeline: Sparse landmarks calibrate dense predictions, which then support cross-view verification for global consistency.

Datasets

The authors evaluated HoloRecon on several challenging benchmarks:

  • 360LOC: A dataset featuring indoor and outdoor scenes with symmetric layouts.
  • YUTO MMS: A massive vehicle-mounted dataset covering over 20 km of urban environment.
  • PanoRecon (New): The authors introduced this new benchmark specifically to test global consistency. it includes seven real-world scenes with repetitive architecture and long-range loops, accompanied by high-accuracy LiDAR ground truth.
Examples from the PanoRecon benchmark, showing the 1.2km and 1.6km routes used for evaluation.

Key Insights and Ideas

  • Geometry as a Filter: The "aha moment" is that predicted local geometry is a powerful tool for disambiguating visual appearance. Two buildings might look the same in a 2D image, but their 3D structure relative to the camera path provides a unique signature that prevents false loop closures.
  • Mutual Dependence: The relationship is symbiotic. The sparse map provides the "skeleton" (global poses and scale) that the dense model needs to stay grounded, while the dense model provides the "flesh" (3D geometry) that allows the sparse map to verify long-range connections.
  • Decoupled Optimization: By caching the dense geometry in local camera coordinates and only updating the transformation matrices, the system achieves global consistency without the prohibitive cost of re-inferring depth for every frame during a loop closure.

Key Results

HoloRecon significantly outperforms existing state-of-the-art methods like COLMAP and Stella-vSLAM, especially on long-range sequences:

  • Trajectory Accuracy: On the YUTO MMS dataset, HoloRecon achieved a mean Absolute Trajectory Error (ATE) of 3.527m, compared to over 68m for the next best SLAM competitor.
  • Reconstruction Fidelity: It achieved the highest F1-scores for dense reconstruction across all benchmarks, particularly excelling in "weakly textured" areas (like plain white walls) where traditional photometric matching usually fails.
  • Efficiency: The GPU-accelerated bundle adjustment backend is roughly 4.3x faster than standard CPU-based implementations, enabling the system to handle the large observation graphs generated by kilometer-long sequences.
Qualitative comparison showing how HoloRecon (right) maintains a clean, single-layer reconstruction compared to the drifted, multi-layer results of other methods.

Limitations and Future Work

While HoloRecon is a major step forward, the authors note a few remaining hurdles:

  • Dynamic Scenes: The current system does not explicitly handle moving objects like cars or pedestrians, which can interfere with feature tracking and depth calibration.
  • Textureless/Blurry Input: Extreme motion blur or very long stretches of featureless walls can still cause the sparse tracking frontend to fail.
  • Real-time Performance: While much faster than SfM, the system is not yet "real-time" due to the overhead of the feed-forward depth prediction.

Future work will likely focus on making the feed-forward models more efficient for streaming video and improving robustness in highly dynamic urban environments.

Summary updated

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/
BibTeX
@misc{li2026holorecon,
  author = {Zhuoxiao Li and Xinyi Liu and Hongye Hou and Jing OU and Kang Du and Shuai Zhang and Dongli Wu and Taoyu Wu and Siyuan Feng and Xiaohui Zhu and Wufan Zhao},
  title = {{HoloRecon: Incremental Reconstruction of Long-Range Panoramas with Global Consistency}},
  year = {2026},
  eprint = {2610.36876},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.36876},
}
Figure