The clearest gain came on aerial blocks
GeoFF3D, a system for building three-dimensional scenes from large UAV image collections, reported better average results than Pi3X + SLRF on nine aerial mapping blocks. Its average Accuracy was 2.72 metres and Completeness was 2.59 metres, while its F@1 and F@5 scores were 0.267 and 0.877. Pi3X + SLRF's F@5 was 0.829.
Relative to Pi3X + SLRF, GeoFF3D's reported Accuracy and Completeness were lower by 18.7% and 15.8%. Its F@1 and F@5 were higher by 0.036 and 0.048. The nine blocks comprised three UseGeo blocks and six UAVFF3D-Real blocks.
A coordinate system built into the model
The model is built around georeferenced camera translations, which tie each camera's estimated position to a world coordinate system, plus optional geometric priors. It predicts camera poses and dense point maps directly in a gravity-aligned, Z-up metric frame.
SLRF, the spatial large-scale reconstruction framework paired with the model, wraps its bounded-view design with footprint-guided chunking, center-outward processing, depth-prior propagation through shared views and hierarchical aggregation. Each leaf chunk combines core views with seam views selected from neighboring regions within a maximum view budget.
The evaluation used a maximum image side length of 518 pixels. GeoFF3D, VGGT + SLRF and Pi3X + SLRF used a 30-view chunk budget, while GeoFF3D and Pi3X + SLRF shared the same perturbed pose priors. The reported geometric scores used shared global Sim(3) and ICP alignment.
The gap was larger on long sequences
On the longer UAVScenes sequences, GeoFF3D reported average Accuracy of 4.14 metres, Completeness of 2.28 metres, F@1 of 0.319 and F@5 of 0.848. Pi3X + SLRF recorded 6.05 metres and 4.40 metres for Accuracy and Completeness, with F@1 of 0.181 and F@5 of 0.687. The reported reductions were 31.5% for Accuracy and 48.2% for Completeness. LingBot-Map's F@5 was 0.451.
The test used eight UAVScenes long sequences sampled with stride 3. It formed a separate long-sequence evaluation from the three UseGeo and six UAVFF3D-Real blocks used for the aerial-block comparison.
Scale came with a chunk-size trade-off
Runtime grew nearly linearly as the number of images increased. In the reported scaling test, GeoFF3D reconstructed 2,000 images in about five minutes while using roughly 16 GiB of GPU memory.
The chunk budget exposed a quality and efficiency trade-off. Twenty views was the fastest setting and used the least memory. Thirty views produced the best reported CD and ATE results with moderate resource use, while 40 views increased peak memory and degraded both global metrics. The main comparison used the 30-view setting.
The component tests were less one-sided
Removing the world-frame losses produced the largest degradation in the reported results. Gravity supervision was associated with lower gravity-direction error, while Sim(3) produced lower ATE but higher CD and gravity-direction error than GA-Sim.
SLRF's pipeline comparison split the advantages across measures. Replacing a temporal-sequential pipeline with a footprint-hierarchical one substantially improved global accuracy. Another variant had the lowest CD, a different variant had the strongest seam consistency, and the full model had the lowest ATE.
The test set had a clear boundary
Under doubled pose noise, GeoFF3D remained relatively stable in the reported robustness test. Using translation priors alone was associated with substantially larger degradation in global accuracy and seam consistency. The reported comparison favored retaining rotation priors alongside translations.
NPU-DroneMap provided a qualitative check because complete dense reference geometry was unavailable. The paper used all 12 keyframe sequences for qualitative evaluation and showed 11 representative sequences without quantitative scoring.
Still a preprint
The document is an arXiv preprint identified as arXiv:2608.28288v2 and dated 31 August 2026. It states that code is available at https://github.com/yanxian-ll/GeoFF3D.
The supplied front matter reports author affiliations and a corresponding-author contact but no funding or conflict-of-interest statement.
Paper data and sources
Original title: GeoFF3D: Coordinate-Anchored Feed-Forward Reconstruction for Large-Scale UAV Mapping
Authors: Xiang Yang, Yongli Wang, Yunsheng Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text