Source-linked AI summary
XDG: Accelerated Visual Disambiguation
Gonglin Chen, Ben Southall, Hanyuan Xiao, Wenbin Teng, Haolin Xiong, Tianwen Fu, Junyi Ouyang, Kshitij Singh Minhas, Supun Samarasekera, Rakesh Kumar, Yajie Zhao
TL;DR
Visual aliasing causes false SfM matches between physically distinct but visually similar surfaces, while prior geometry-aware disambiguation remains expensive because of its heavy transformer classifier. XDG adapts Depth Anything 3 with LoRA and classifies from compact camera tokens using a lightweight MLP. It remains competitive with DG++ across pairwise and reconstruction benchmarks while reducing disambiguation time by more than 3× and saving over 10 hours on individual LaMAR scenes.
Problem
Visual aliasing allows distinct surfaces to create false SfM matches, and existing geometry-aware disambiguation uses a heavy transformer classifier that makes large-scale processing expensive.
Method
XDG fine-tunes Depth Anything 3 with LoRA adapters, repurposes its camera tokens as pair-level representations, and predicts matches with a lightweight MLP.
Results
More than 3× lower disambiguation time accompanies performance comparable to DG++ across pairwise and downstream reconstruction tasks.
Takeaways & Limitations
XDG substantially reduces visual-disambiguation cost while preserving competitive reconstruction accuracy for large-scale 3D reconstruction.
Takeaways & Limitations
Local matching remains vulnerable to doppelganger patterns, and AerialMegaDepth poses are approximate rather than fully reliable because images are registered to geospatial meshes.
Abstract
from arXiv · showhide
Visual aliasing, also known as the doppelganger problem, remains a key challenge for structure-from-motion (SfM): visually similar but physically distinct surfaces can produce incorrect image matches and degrade reconstruction quality. Previous work mitigates this issue with geometry-aware foundation-model features, but places a heavy transformer classifier on top of the backbone, making large-scale disambiguation expensive. We introduce XDG, an efficient visual disambiguation model designed for scalable SfM. Our key observation is that a 3D foundation model already performs the cross-view geometric reasoning necessary for visual disambiguation, so doppelganger classification should adapt the backbone representation directly rather than relearn pair reasoning in a separate heavy decoder. XDG fine-tunes Depth Anything 3 with lightweight LoRA adapters and repurposes its camera tokens as compact pair-level classification tokens. A compact MLP head predicts whether a candidate image pair observes the same 3D surface. Extensive experiments show that XDG provides a favorable accuracy-efficiency tradeoff: it remains competitive with the state-of-the-art disambiguation method across pairwise and reconstruction benchmarks and delivers more than a 3x inference speedup. On individual LaMAR scenes containing thousands of images, XDG saves more than 10 hours of visual disambiguation processing. Code is available at https://github.com/xtcpete/xdg.
1. Introduction
XDG addresses visual aliasing in SfM by adapting a geometry-aware 3D backbone directly, replacing heavy pairwise reasoning with compact camera-token classification. It achieves competitive disambiguation and reconstruction performance while substantially reducing processing cost.
- Visual aliasing lets distinct physical surfaces produce false image matches that contaminate SfM reconstructions.Such errors can merge unrelated structures, misregister cameras, or fragment reconstructions.
- XDG adapts Depth Anything 3 with parameter-efficient LoRA modules and classifies pairs from its native stage-wise camera tokens.The camera tokens already summarize cross-view relationships through DA3’s local and global feature interactions.
- XDG removes the heavy post-backbone transformer, fuses representations from both image orders, and uses a lightweight MLP for final prediction.The design enforces order consistency through symmetric token aggregation while keeping classification compact.
- More than 3× lower per-pair inference time accompanies accuracy comparable to DG++ on pairwise benchmarks and competitive performance on reconstruction benchmarks.The evaluation includes DG, VisymScenes, WRIVA, and LaMAR.
- 10.85–13.91 hours of visual-disambiguation processing are saved per LaMAR scene containing 7.5K–9.3K images.The comparison is against DG++ and illustrates the scaling benefit on large image collections.
2. Related Work
Prior SfM and image-matching methods improve correspondence estimation but remain vulnerable to visually aliased surfaces and expensive ambiguity reasoning. XDG follows geometry-aware disambiguation while reducing cost by adapting DA3 and classifying directly from camera tokens.
- Local feature matchers improve correspondence robustness but can produce spurious matches between distinct yet visually similar surfaces.Their patch- or keypoint-level operation does not explicitly model global image context or higher-level 3D scene consistency.
- Earlier 3D learning methods use feed-forward neural networks to recover scene geometry and camera poses from uncalibrated images.The passage situates XDG within the broader shift from keypoint matching and iterative optimization toward feed-forward reconstruction.
- XDG complements hybrid reconstruction by replacing DG++ with a faster classifier while preserving competitive reconstruction accuracy in most settings.This route targets the cost of the two-view disambiguation stage rather than replacing the entire reconstruction framework.
- SfM geometric verification and robust optimization can still admit incorrect edges from repeated structures and near-duplicate appearances.Visual disambiguation must determine whether plausible correspondences arise from the same 3D structure.
- XDG processes both image orders, aggregates their camera tokens consistently, and avoids an expensive transformer reasoning module after the backbone.This is the method’s efficiency route for large collections requiring many candidate-edge evaluations.
3. Method
XDG performs SfM doppelganger classification by adapting DA3 directly and classifying compact camera-token representations instead of using a heavy post-backbone transformer. It combines LoRA adaptation, symmetric two-order token aggregation, and a lightweight MLP head for efficient pair prediction.
- Model Architecture: XDG formulates candidate-pair disambiguation as binary classification of true matches versus doppelganger edges before reconstruction.The classifier determines whether two images observe a common 3D surface or should be pruned from the SfM match graph.
- Model Architecture: XDG extracts DA3 camera tokens from both image orders, realigns reversed outputs, and fuses corresponding tokens into an order-consistent pair representation.A shared two-layer fusion MLP maps concatenated token features from 2Ce to 4Ce and back to Ce at each stage.
- Model Architecture: XDG uses frozen DA3 with LoRA adapters, while leaving camera-token parameters trainable and adapting selected attention and feed-forward projections.LoRA adds a trainable low-rank residual to the pretrained linear projections; the implementation uses rank r = 8 and α = 16.
- Model Architecture: XDG classifies pairs using only compact camera tokens rather than passing dense tokens to a heavy transformer classifier.The method’s efficiency strategy is to reuse the backbone’s native cross-view representation instead of adding a separate transformer reasoning module.
- Model Architecture: The classification head projects fused stage tokens, applies layer normalization, averages across four stages and two views, and predicts two class logits with a compact MLP.The logits represent the doppelganger and true-match classes; the head uses Cd = Ch = 768 and dropout 0.1.
- Implementation Details: XDG is trained with focal loss and randomized order and resolution augmentation using labeled pairs from DG and VisymScenes.Training runs for 10 epochs with AdamW; on 8 NVIDIA H100 GPUs, the model converges in 4 hours.
4. Experiments
XDG achieves competitive visual disambiguation and reconstruction quality while substantially reducing inference and total processing time across training-domain and out-of-domain benchmarks.
- Pairwise visual disambiguation: XDG achieves 0.978 AP and 0.975 ROC AUC on DG, approaching DG++ while improving recall at 0.99 precision from 0.642 to 0.702.On VisymScenes, XDG slightly outperforms DG++ in AP, ROC AUC, and recall at 0.99 precision.
- Pairwise visual disambiguation: XDG reduces per-pair runtime from 118.3 ms to 34.2 ms on DG and from 115.1 ms to 33.5 ms on VisymScenes, exceeding 3× speedup over DG++.It also runs more than 5× faster than DG-OG while maintaining competitive pairwise accuracy.
- Training-domain reconstruction: On AerialMegaDepth, XDG and DG++ achieve nearly identical pose accuracy while XDG reduces total disambiguation time from 5.28 to 1.20 hours, saving 4.08 hours.The benchmark averages pose AUC across 8 scenes using real images as reconstruction inputs.
- Training-domain reconstruction: Across five VisymScenes scenes, XDG registers similar image counts and inlier ratios to DG++ while reducing disambiguation time from 1.20 hours to 0.32 hours, a 3.7× speedup.Qualitative reconstructions show XDG removes incorrectly registered cameras and produces layouts visually comparable to DG++.
- Out-of-domain reconstruction: On WRIVA, both methods improve reconstruction over no disambiguation; DG++ is slightly better at the tightest thresholds, while XDG achieves slightly higher AUC@10° and AUC@30°.XDG reduces visual-disambiguation time from 9.86 to 3.06 hours across 34 sequences, saving 6.80 hours.
- Out-of-domain reconstruction: On LaMAR, DG++ generalizes better on indoor CAB, while XDG is comparable outdoors and reduces total disambiguation time from 49.93 to 15.28 hours.XDG saves 10.85 hours on HGE and 13.91 hours on LIN, with a 3.27× speedup across all three scenes.
5. Conclusion
XDG targets the runtime bottleneck of large-scale visual disambiguation while preserving reconstruction performance comparable to DG++. Its speedups accumulate substantially on large benchmarks.
- XDG achieves performance comparable to DG++ across pairwise and downstream reconstruction tasks while reducing disambiguation time by more than 3×.The model adapts a 3D foundation model with LoRA and replaces heavy transformer classifiers with lightweight token aggregation and an MLP head.
- 34.65 hours are saved over LaMAR, and more than 6 hours are saved on WRIVA.These savings reflect visual-disambiguation processing time.
- XDG makes visual disambiguation significantly more efficient while preserving reconstruction accuracy.
A. Additional Reconstruction Results
Additional experiments show that visual disambiguation interacts strongly with the reconstruction pipeline and dataset difficulty. XDG and DG++ are often comparable, but filtering can reduce accuracy when correspondence formation is already limited.
- A.1. WRIVA: On WRIVA, both DG++ and XDG reduce pose AUC relative to vanilla COLMAP.The result is averaged across 34 WRIVA sequences.
- A.1. WRIVA: WRIVA’s viewpoint, altitude, and appearance variation can leave too few tracks after disambiguation prunes an already sparse match graph.This suggests correspondence formation, rather than false-edge removal alone, is the primary COLMAP bottleneck on WRIVA.
- A.2. AerialMegaDepth: Both DG++ and XDG substantially improve pose AUC over reconstruction without visual disambiguation on GLUEMAP AerialMegaDepth, with comparable performance across thresholds.
- A.2. AerialMegaDepth: GLUEMAP’s lower results reflect difficult real images that can introduce inaccurate camera estimates when all input images are incorporated.Despite this, both visual-disambiguation methods remain beneficial in this reconstruction setting.
B. More Ablation Results
Ablations identify the components that balance XDG’s accuracy and efficiency. Bidirectional processing and training diversity improve robustness, while backbone capacity controls the latency–accuracy tradeoff.
- Cumulative ablation: Feature aggregation, the MLP head, and camera-token classification reduce inference cost, while LoRA adapts the backbone to recover task accuracy.
- Backbone capacity: DA3-Base provides the best overall accuracy–efficiency compromise among the tested backbone capacities.DA3-Small roughly halves latency but reduces accuracy, while DA3-Large increases latency by approximately 2.4× over DA3-Base.
- Training strategy: Removing random input-order flipping or multi-resolution sampling consistently lowers AP and ROC AUC on both test sets.The effect of removing multi-resolution training is particularly visible on VisymScenes.
- Order robustness: Single-order processing reduces latency from 34.2 to 19.7 ms on DG and from 33.5 to 19.6 ms on VisymScenes, but lowers AP and ROC AUC.
C.1. AerialMegaDepth: Per-Scene Results
Per-scene AerialMegaDepth results show that XDG’s gains are concentrated on scenes with severe ambiguity from repeated structures. Bidirectional aggregation also removes dependence on arbitrary image ordering.
- Per-scene reconstruction: XDG’s gains are concentrated on scenes with severe ambiguity, notably Cologne Cathedral and the Louvre Museum.On easier scenes, all three configurations are similar.
- Per-scene reconstruction: Across AerialMegaDepth scenes, XDG closely tracks DG++ while requiring substantially less visual-disambiguation time.
- Order robustness: Bidirectional XDG produces the same matching confidence after swapping image order, unlike the order-sensitive single-order model.At τ = 0.8, the single-order model rejects A→B but accepts B→A.
C.2. WRIVA: Per-Sequence Results
Per-sequence WRIVA results vary substantially by capture condition, with visual disambiguation helping many difficult sequences but not every small or already well-reconstructed case.
- Visual disambiguation provides large reconstruction gains on many difficult PTZ and varying-altitude sequences.
- A few small or already well-reconstructed sequences favor using no filtering.
- XDG and DG++ show similar overall behavior while exhibiting complementary strengths across individual sequences.
D. More Qualitative Results
Additional visualizations compare ground-truth camera layouts with reconstructions using no disambiguation, DG++, and XDG across AerialMegaDepth and WRIVA examples.
- AerialMegaDepth: Figures 6–8 compare ground-truth and reconstructed camera layouts for six AerialMegaDepth scenes.
- WRIVA: Figures 9–11 present six WRIVA GLUEMAP reconstructions selected where DG++ and XDG outperform reconstruction without visual disambiguation.
- WRIVA: Across the selected WRIVA examples, disambiguation removes geometrically inconsistent matches and produces camera layouts that better agree with ground truth.
- AerialMegaDepth: For Ponte di Rialto, DG++ and XDG each produce two reconstruction components, shown stacked vertically.
- WRIVA: On WRIVA Sequence 00, AUC@10° rises from 2.63% without disambiguation to 38.38% with DG++ and 38.45% with XDG.
- WRIVA: On WRIVA Varying Altitudes Multi-Zone, AUC@10° rises from 25.41% without disambiguation to 52.25% with DG++ and 69.93% with XDG.