PaperlayerView exact revision on arXiv

Source-linked AI summary

GeoFF3D: Coordinate-Anchored Feed-Forward Reconstruction for Large-Scale UAV Mapping

Xiang Yang, Yongli Wang, Yunsheng Zhang

arXiv:2608.28288v1cs.CV

TL;DR

Large-scale UAV mapping requires feed-forward reconstruction beyond bounded image sets, with unified georeferenced outputs and stable aggregation despite near-collinear trajectories. GeoFF3D combines coordinate-anchored prediction in a gravity-aligned Z-up metric frame with SLRF for spatial chunking and hierarchical aggregation. It achieves the best average performance on aerial blocks, stronger results on long UAV sequences, and reconstructs 2,000 images in approximately five minutes.

  • Problem

    Existing feed-forward reconstruction methods process bounded image sets in local or internally normalized frames, while full Sim(3) alignment may become unstable for near-collinear UAV trajectories.

  • Method

    GeoFF3D uses georeferenced translations and optional geometric priors for coordinate-anchored prediction, while SLRF performs spatial chunking, prior propagation, and gravity-preserving hierarchical aggregation.

  • Results

    GeoFF3D achieves the best average performance on aerial mapping blocks and a larger advantage on long, near-collinear sequences, while reconstructing 2,000 images in about five minutes.

  • Takeaways & Limitations

    Coordinate anchoring benefits general multi-strip aggregation, with metric scale and gravity especially important for elongated trajectories.

Abstract

from arXiv · show

Existing feed-forward 3D reconstruction methods typically process a bounded number of images and recover cameras and geometry in local or internally normalized frames. Extending them to large-scale UAV mapping requires scalable multi-chunk processing and reliable aggregation, while full Sim(3) alignment can become unstable for near collinear trajectories. We present GeoFF3D, which combines a coordinate-anchored model with a spatial large-scale reconstruction framework (SLRF). The model uses georeferenced camera translations and optional geometric priors to predict camera poses and dense point maps directly in a gravity-aligned Z-up metric frame. SLRF partitions images into spatially overlapping chunks, propagates shared-view priors, and aggregates local reconstructions hierarchically, while remaining applicable to different bounded-view models. Across nine aerial mapping blocks, GeoFF3D achieves the best average reconstruction quality, improving F@5 from 0.829 for Pi3X + SLRF to 0.877. On long UAVScenes sequences, it reaches 0.848, compared with 0.687 for Pi3X + SLRF and 0.451 for the strongest evaluated SLAM/streaming baseline. GeoFF3D reconstructs 2,000 images in approximately five minutes, demonstrating scalable and robust large-scale UAV reconstruction.The code is available at https://github.com/yanxian-ll/GeoFF3D.

1 Introduction

GeoFF3D addresses large-scale UAV reconstruction by anchoring predictions directly in a georeferenced gravity-aligned frame and aggregating spatially overlapping chunks. It achieves strong performance across aerial blocks and long UAV sequences while scaling to thousands of images.

  • Motivation: Near-collinear UAV trajectories can make full Sim(3) alignment unstable, weakening roll and pitch constraints and tilting dense reconstructions.
  • Method: GeoFF3D defines the prediction frame using georeferencing priors, directly producing camera poses and dense point maps in a gravity-aligned Z-up metric frame.
  • Method: SLRF uses spatially overlapping chunks, center-outward processing, shared-view prior propagation, and gravity-preserving hierarchical aggregation to improve cross-chunk consistency.
  • Results: GeoFF3D achieves the best average performance across nine aerial blocks and a larger advantage on long, near-collinear UAV sequences.
  • Results: GeoFF3D reconstructs 2,000 images in about five minutes, demonstrating scalability to large UAV collections.

2 Related Work

Related work spans bounded-view feed-forward reconstruction and scalable systems for large image collections. Existing approaches use internal normalization, spatial or streaming mechanisms, and global consistency procedures such as registration, loop closure, pose graphs, or geometric optimization.

  • Feed-forward 3D reconstruction: DUSt3R, MASt3R, and VGGT anchor predictions to a selected input camera, while π3 predicts permutation-equivariant outputs in an internally normalized representation.
  • Scalable feed-forward reconstruction: Scalable feed-forward methods extend reconstruction through spatial memories, persistent states, causal attention, token compression, or test-time adaptation.
  • Scalable feed-forward reconstruction: Other systems reconstruct overlapping clips or submaps and enforce global consistency through learned registration, loop closure, pose graphs, or geometric optimization.

3 Method

GeoFF3D reconstructs cameras and dense geometry directly in a shared georeferenced Z-up metric frame, then scales bounded-view inference through spatial chunking, prior propagation, and hierarchical gravity-preserving aggregation.

  • Coordinate-Anchored Model: GeoFF3D predicts camera poses and dense point maps in a shared georeferenced Z-up metric frame using georeferencing inputs and optional geometric priors.The coordinate-anchored model operates on bounded-view spatial chunks and restores predictions to the georeferenced metric frame after normalization.
  • Coordinate-Anchored Model: The model fuses image, ray, depth, and georeferenced view features with prior-availability masks, then predicts cameras, dense geometry, and per-pixel confidence.Prior dropout trains the model to handle complete, partial, or sparse geometric observations, while confidence suppresses unreliable geometry during fusion.
  • Training: Training combines local, world-frame, pose, and gravity losses, followed by high-resolution fine-tuning with randomly dropped and perturbed geometric priors.The model is initialized from Pi3X and trained on UAVFF3D and BlendedMVS.
  • Spatial Large-Scale Reconstruction Framework: SLRF partitions large image collections into spatially compact, overlapping chunks and processes them from the scene center outward.Each image belongs to one core set and may also serve as a seam view for adjacent chunks, preserving overlap for propagation and alignment.
  • Spatial Large-Scale Reconstruction Framework: Shared-view depth priors are cached between chunks, while low-confidence regions are filtered before updating the cache and continuing inference.The framework uses a confidence quantile threshold of 0.25 for removing unreliable regions.
  • Gravity-Aligned Hierarchical Aggregation: Leaf chunks are anchored with gravity-aligned similarity transformations, sibling nodes are refined with gravity-aligned rigid transformations, and the tree is merged bottom-up.GA-Sim estimates scale, yaw, and translation while preserving roll and pitch; GA-Rigid estimates residual yaw and translation from shared-view correspondences.

4 Experiments

GeoFF3D is evaluated on aerial blocks and long UAV sequences against feed-forward baselines using common metrics and reconstruction protocols. It achieves the strongest average quality, remains robust in ablations, and scales to thousands of images.

  • Experimental setup: GeoFF3D is evaluated on nine aerial blocks and eight UAVScenes sequences, with additional qualitative NPU-DroneMap sequences.The aerial blocks contain 224–1,177 images per scene, while sampled UAVScenes sequences contain 1,317–2,589 original frames.
  • Experimental setup: Aerial-block evaluation reports Accuracy, Completeness, F@1, and F@5, with scene averages and separate global alignment for geometry evaluation.F@1 and F@5 are measured at 1 m and 5 m thresholds, respectively.
  • Large-scale UAV reconstruction: GeoFF3D achieves the best average aerial-block results: Accuracy 2.72 m, Completeness 2.59 m, F@1 0.267, and F@5 0.877.Compared with Pi3X + SLRF, it improves F@1 by 0.036 and F@5 by 0.048.
  • Large-scale UAV reconstruction: On long UAVScenes sequences, GeoFF3D improves F@5 from 0.687 for Pi3X + SLRF to 0.848 and reduces Accuracy from 6.05 m to 4.14 m.Completeness decreases from 4.40 m to 2.28 m, while F@1 increases from 0.181 to 0.319.
  • Ablations and analysis: Ablations show that world-frame and gravity losses support coordinate anchoring, while GA-Sim preserves gravity-aligned geometry better than Sim(3) despite higher ATE.The SLRF variants also differ in global accuracy and seam consistency, and pose-prior removal or increased noise degrades performance.
  • Ablations and analysis: A 30-view chunk budget balances reconstruction quality and efficiency, while runtime grows nearly linearly and 2,000 images require about five minutes and roughly 16 GiB of GPU memory.Twenty views are fastest and use least memory; forty views increase memory and degrade global metrics.

5 Conclusions

GeoFF3D reconstructs cameras and dense geometry directly in a georeferenced, gravity-aligned Z-up metric frame. It performs best on aerial blocks, gains more on long near-collinear sequences, and supports large-scale aggregation.

  • GeoFF3D directly reconstructs cameras and dense geometry in a georeferenced, gravity-aligned Z-up metric frame.
  • GeoFF3D achieves the best average performance on aerial mapping blocks and a larger advantage on long, near-collinear sequences.
  • The results identify metric scale and gravity as especially important for elongated trajectories.

A Per-Sequence Quantitative Results

The per-sequence analysis exposes variation hidden by UAVScenes averages by separating directed distance errors from thresholded F-scores.

  • Per-sequence results reveal variation that is hidden by main-paper UAVScenes averages.
  • The analysis separates directed distance errors from thresholded F-scores.

A.1 UAVScenes

The UAVScenes appendix reports per-sequence distance and thresholded F-score results under the main-paper evaluation configuration.

  • Evaluation configuration: VGGT-SLAM uses Sim(3) alignment, all SLRF methods use 30-view chunks, and Pi3X + SLRF and GeoFF3D receive identical noisy pose priors.
  • Distance metrics: Table 7 reports per-sequence Accuracy and Completeness on UAVScenes, with both metrics measured in meters.
  • Thresholded metrics: Table 8 reports per-sequence F-scores on UAVScenes at 1 m and 5 m thresholds.

B Qualitative Results

The qualitative results complement aggregate metrics by making reconstruction coverage, continuity, orientation, and generalization visually assessable. Matched-camera views and explicit acquisition geometry support direct comparison across challenging aerial blocks.

  • B Qualitative Results: Qualitative examples assess coverage, boundary continuity, vertical orientation, and cross-dataset generalization.The examples are organized by aerial blocks, long trajectories, and NPU-DroneMap sequences.
  • B Qualitative Results: Matched cameras within each row enable direct comparison of coverage, boundary completeness, local deformation, and missing regions.Trajectory views also expose the acquisition geometry of difficult near-collinear or sparsely overlapping blocks.
  • B Qualitative Results: Trajectory columns make acquisition geometry explicit when interpreting difficult near-collinear or sparsely overlapping blocks.

VGGT + SLRF

Figures 7 and 8 present matched-view qualitative comparisons for GeoFF3D, Pi3X + SLRF, and VGGT + SLRF across aerial mapping blocks. Consistent cameras and spatial bounds make visual differences directly comparable.

  • VGGT + SLRF: Figure 7 compares trajectory, ground truth, GeoFF3D, Pi3X + SLRF, and VGGT + SLRF across UseGeo blocks D1–D3.All reconstructions in each row use the same rendering camera and bounds.
  • VGGT + SLRF: Figure 8 compares the same reconstruction views across five UAVFF3D-Real blocks: NF01, XX01, YHT01, NF02, and XX02.Each row uses the same view and spatial bounds for comparison.

B.2 Long UAV Sequences

Long, near-collinear trajectories are evaluated with both top views and vertical side profiles because horizontal layout alone may conceal incorrect vertical orientation or scene thickness. GeoFF3D preserves a gravity-constrained Z-up profile during hierarchical aggregation.

  • B.2 Long UAV Sequences: Near-collinear long trajectories are sensitive to post-hoc roll and pitch estimation, motivating top-view and vertical side-profile comparisons.
  • B.2 Long UAV Sequences: GeoFF3D retains a Z-up profile because gravity is constrained during prediction and preserved during hierarchical aggregation.
  • B.2 Long UAV Sequences: Figure 9 compares representative Town01, Valley01, Airport01, and Island01 trajectories across GeoFF3D, chunked or streaming baselines, and LingBot-Map.Shared views expose fragmentation and global orientation errors that aggregate metrics can obscure.

B.3 NPU-DroneMap Generalization

NPU-DroneMap provides real keyframe sequences across varied environments for testing generalization beyond quantitative benchmark datasets. Figure 10 covers representative sequences without dense reference geometry for quantitative scoring.

  • B.3 NPU-DroneMap Generalization: NPU-DroneMap contains real keyframe sequences spanning urban, industrial, road, and agricultural environments.
  • B.3 NPU-DroneMap Generalization: These sequences test whether the coordinate-anchored model and spatial pipeline generalize beyond quantitative benchmark datasets.
  • B.3 NPU-DroneMap Generalization: Figure 10 covers 11 representative NPU-DroneMap keyframe sequences across campus, factory, road, industrial, village, and agricultural scenes.No dense reference geometry is available for quantitative scoring.
Loading 2608.28288v1…