Edge Prioritization for Robust SfM

• Edge Prioritization is a strategy that optimizes pose graph initialization in SfM by ranking candidate image pairs based on learned geometric reliability.
• It combines a graph neural network predictor with multi-MST selection to fuse local descriptor similarity and global connectivity, enhancing overall graph structure.
• The method reduces expensive pairwise verifications while improving downstream motion averaging and bundle adjustment through informed and sparse edge selection.

Edge prioritization is a globally informed strategy for pose graph initialization in Structure-from-Motion (SfM). In this setting, images are nodes, candidate image pairs are edges, and initialization determines which pairs will be sent to expensive geometric verification. Rather than connecting each image only to its $k$ nearest neighbors in descriptor space, edge prioritization ranks candidate edges by their utility for downstream geometry and constructs a sparse but well-connected pose graph by combining a learned reliability predictor, multi-minimum-spanning-tree (MST) selection, and connectivity-aware score modulation. The resulting initialization is designed to produce more reliable and compact pose graphs, particularly in sparse, high-speed, and ambiguous-scene regimes (Wei et al., 25 Feb 2026).

1. Formal setting and motivation

Let $G=(V,E)$ be a graph over $N$ images, with $V=\{I_1,\dots,I_N\}$. A candidate undirected edge $e_{ij}\in E$ denotes a potential relative pose $T_{ij}\in SE(3)$ between images $I_i$ and $I_j$. Edge prioritization assigns each candidate pair a utility or reliability score $r(e_{ij})\in[0,1]$, abbreviated $r_{ij}$, intended to measure how useful that pair is for accurate and stable SfM. The pose-graph structure is further characterized by hop-count shortest-path distance $G=(V,E)$0, graph diameter $G=(V,E)$1, and normalized distance $G=(V,E)$2 when $G=(V,E)$3, or $G=(V,E)$4 otherwise (Wei et al., 25 Feb 2026).

The method is motivated by limitations of standard image-retrieval-based initialization. Conventional pipelines connect each image to its $G=(V,E)$5 nearest neighbors in descriptor space, making independent local decisions that ignore the global structure of the image set. The stated failure modes are long chains with large diameter, which make global pose estimation fragile and sensitive to drift, and fragmentation into weakly coupled substructures that are rarely repaired later. The computational pressure is also explicit: verifying all $G=(V,E)$6 pairs is infeasible, and if $G=(V,E)$7 edges are selected, verification cost scales roughly linearly in $G=(V,E)$8 and in the per-pair RANSAC iterations, which can be large. This motivates ranking edges not only by local similarity but also by their contribution to global consistency, including loop constraints such as $G=(V,E)$9.

In this sense, edge prioritization is not merely a retrieval heuristic. It is a graph-construction policy that attempts to allocate the verification budget to edges that both match well and strengthen graph topology. A plausible implication is that initialization quality becomes a first-order determinant of later motion averaging and bundle adjustment conditioning, rather than a simple pre-processing step.

2. Learned estimation of edge reliability

The reliability predictor is a graph neural network trained with SfM-derived supervision. Each image $N$0 is first encoded to a descriptor $N$1, with the paper giving DINOv2 with SALAD/MegaLoc as examples. On the complete graph, each directed edge is initialized by fusing the two descriptors and their cosine similarity:

$N$2

Message passing then runs for two iterations. The edge update is

$N$3

neighbor messages are

$N$4

and node states are updated as

$N$5

The final edge reliability prediction is

$N$6

Supervision is self-supervised and derived from a standard SfM pipeline. For each pair $N$7, the first signal is RANSAC verifiability,

$N$8

and the second is multi-view utility,

$N$9

These are combined into a continuous relevance label

$V=\{I_1,\dots,I_N\}$0

where the normalization maps small inlier counts, such as those below $V=\{I_1,\dots,I_N\}$1, into $V=\{I_1,\dots,I_N\}$2 and larger counts into $V=\{I_1,\dots,I_N\}$3 to avoid domination by extremely easy pairs.

Training optimizes edge ordering within each scene rather than raw regression. The ranking objective is based on DCG and NDCG:

$V=\{I_1,\dots,I_N\}$4

with NDCGLoss2++ (LambdaLoss) as the differentiable surrogate. The only stated hyperparameter is the sampled list size $V=\{I_1,\dots,I_N\}$5, set to half the list per training instance. This formulation emphasizes relative ordering of candidate edges, which is directly aligned with the downstream selection problem.

3. Multi-MST selection and connectivity-aware modulation

After prediction, edge prioritization converts scores into a sparse pose graph through repeated MST construction. At iteration $V=\{I_1,\dots,I_N\}$6, let $V=\{I_1,\dots,I_N\}$7 be the union of previously selected trees. Hop-count distances $V=\{I_1,\dots,I_N\}$8 and diameter $V=\{I_1,\dots,I_N\}$9 are computed, and normalized distances are defined by

$e_{ij}\in E$0

These distances modulate the learned reliability:

$e_{ij}\in E$1

The stated intuition is that edges already inside dense regions retain their local reliability, whereas edges bridging long chains receive a boost. This suggests a deliberate trade-off between local matchability and graph-wide compactness.

The modulation is not applied indiscriminately. Only the top-5 candidate edges per image, ranked by $e_{ij}\in E$2, are updated, and edges with $e_{ij}\in E$3 are discarded from modulation to avoid boosting weak pairs. Previously selected edges are masked by assigning score $e_{ij}\in E$4 before the next MST. These safeguards are part of the method rather than auxiliary heuristics.

Scores are converted to weights through

$e_{ij}\in E$5

where $e_{ij}\in E$6 at iteration $e_{ij}\in E$7, and for $e_{ij}\in E$8, $e_{ij}\in E$9. For each $T_{ij}\in SE(3)$0, an MST $T_{ij}\in SE(3)$1 is computed with Kruskal’s algorithm, then its edges are masked so the next tree must diversify. The final initialized pose graph is

$T_{ij}\in SE(3)$2

Because each MST contributes $T_{ij}\in SE(3)$3 edges, the union remains sparse while introducing redundancy and complementary long-range links. In graph-theoretic terms, the construction acts as a sparsifier that preserves connectivity and reduces diameter without reverting to dense verification.

4. Role inside the SfM pipeline and computational profile

The selected edge set $T_{ij}\in SE(3)$4 is passed to feature matching and robust two-view estimation. The paper lists SuperPoint + LightGlue followed by RANSAC, MAGSAC, or VSAC as examples. Verified pairs yield essential or fundamental matrices, relative rotations and translations, and inlier sets, after which incremental or global SfM proceeds with initialization and bundle adjustment. On VisymScenes, RANSAC used $T_{ij}\in SE(3)$5k iterations. Because verification is expensive, prioritization is intended to reduce the number of candidate pairs sent to verification, shorten chains, and improve conditioning for motion averaging and bundle adjustment (Wei et al., 25 Feb 2026).

The main computational burden of the predictor is the complete graph. With $T_{ij}\in SE(3)$6 message-passing iterations, GNN inference has $T_{ij}\in SE(3)$7 compute and $T_{ij}\in SE(3)$8 memory. To manage large $T_{ij}\in SE(3)$9, the method uses graph clustering with METIS: the complete graph is partitioned into subgraphs, the number of clusters is set to $I_i$0, each subgraph is processed with 1-hop expansion, and overlapping predictions are averaged. The purpose is to avoid GPU memory exhaustion while preserving local global context.

MST construction with Kruskal costs $I_i$1 per tree, with $I_i$2 on the complete graph, and scales linearly in the number of trees $I_i$3. Reported IMC23 average runtimes for $I_i$4 MSTs are: encoder $I_i$5 s; predictor $I_i$6 s for the proposed method versus $I_i$7 s for MegaLoc; MST $I_i$8 s versus $I_i$9 s; and COLMAP mapping approximately $I_j$0k s versus $I_j$1k s. The explicit interpretation given is that predictor overhead is negligible relative to downstream SfM, and that improved selection reduces mapping time.

5. Empirical behavior, benchmarks, and ablations

Training used MegaDepth with $I_j$2 scenes and $I_j$3 held out. Evaluation covered MegaDepth scenes “0015” and “0022”, IMC23 PhotoTourism with $I_j$4 scenes, and VisymScenes with $I_j$5 ambiguous scenes containing doppelgangers. Metrics were relative pose AUC@$I_j$6 and AUC@$I_j$7, percentage of registered cameras, mapping time, and for VisymScenes the fraction of correctly localized cameras within a geolocation threshold (Wei et al., 25 Feb 2026).

Across $I_j$8 MSTs, the method is reported to achieve the highest AUC on PhotoTourism and MegaDepth. Gains are largest when $I_j$9 or $r(e_{ij})\in[0,1]$0, where the graph is extremely sparse; methods converge as $r(e_{ij})\in[0,1]$1 increases, but the proposed method remains best. Registration rates on PhotoTourism are near $r(e_{ij})\in[0,1]$2 across methods, so the measured AUC gains are attributed to selecting more informative edges rather than registering more cameras. Runtime-versus-accuracy curves lie on the Pareto frontier, yielding higher AUC for comparable COLMAP time.

The ambiguous-scene results are especially central. On VisymScenes, the method reconstructs the largest fraction of correct cameras across all tested $r(e_{ij})\in[0,1]$3. Connectivity-aware modulation improves AUC@$r(e_{ij})\in[0,1]$4 from $r(e_{ij})\in[0,1]$5, $r(e_{ij})\in[0,1]$6, and $r(e_{ij})\in[0,1]$7 without modulation to $r(e_{ij})\in[0,1]$8, $r(e_{ij})\in[0,1]$9, and $r_{ij}$0 with modulation for $r_{ij}$1, $r_{ij}$2, and $r_{ij}$3, respectively. Applying Doppelgangers++ on top of the predicted ranks produced no further benefit, which the paper interprets as evidence that the predictor already suppresses distractors.

Ablations isolate several components. Multi-MST selection outperforms kNN edge selection across the board. Oracle rankings based on RANSAC inliers or common 3D points indicate that MST selection aligns more closely with ground-truth geometry than kNN. Within the modulation subsystem, enabling modulation, restricting updates to top-5 edges per image, and normalizing distances gives the strongest AUC across $r_{ij}$4. With a weaker SALAD backbone, the method still shows strong improvements, whereas removing the GNN causes a large drop at $r_{ij}$5, highlighting the value of global message passing. An appendix comparison for IMC23 reports kNN AUC@$r_{ij}$6 values of $r_{ij}$7, $r_{ij}$8, $r_{ij}$9, and $G=(V,E)$00 for MegaLoc and $G=(V,E)$01, $G=(V,E)$02, $G=(V,E)$03, and $G=(V,E)$04 for the proposed method at $G=(V,E)$05, $G=(V,E)$06, $G=(V,E)$07, and $G=(V,E)$08.

6. Limitations, related concepts, and broader uses of the term

The stated failure cases are extremely sparse overlap, very low-resolution views with small visible landmark portions, and severe appearance ambiguity combined with weak geometric cues. In such regimes, both retrieval and the GNN may leave too few verifiable edges. The method also assumes that descriptors contain sufficient global context; a poor backbone reduces gains, although the GNN and geometric supervision mitigate this. Recommended defaults are two message-passing iterations, $G=(V,E)$09–$G=(V,E)$10 MSTs for sparse or high-speed settings, up to $G=(V,E)$11 for more redundancy, modulation weight $G=(V,E)$12 in the range $G=(V,E)$13–$G=(V,E)$14, score modulation restricted to top-5 candidates per image, and ignoring edges with $G=(V,E)$15. For scalability, METIS clustering is recommended for $G=(V,E)$16, and for online or SLAM settings the predictor can run in sliding windows with MSTs constructed per window and merged over time (Wei et al., 25 Feb 2026).

Conceptually, the method is linked to cycle consistency and loop constraints, graph sparsification, active selection, and global pose estimation. It does not explicitly optimize cycle-error terms, but it uses inliers and 3D overlap to learn edge reliability and then exploits graph distances to favor loop-closing or chain-shortening links. A plausible implication is that edge prioritization shifts pose-graph initialization from a retrieval-only problem toward a learned global graph design problem.

The phrase “edge prioritization” is also used in other research areas, but with different meanings. In adaptive edge caching it refers to prioritized content admission and eviction under an SMDP/DDQL framework (Niknia et al., 2024). In distributed edge-cloud scheduling it refers to priority-aware task assignment and preemption (Rublein et al., 2024). In SDN/OpenFlow edge networks it denotes prioritized multi-tenant bandwidth allocation through meters and queues (Shahriar et al., 2024). In causal fairness research it refers to prioritizing unfair edges in a causal graph for discrimination mitigation (Ravishankar et al., 2021). These usages are terminologically related but methodologically distinct from pose-graph initialization in SfM.