acceptodds
Under review as a conference paper at ICLR 2027

Translation Averaging as Global Discrete Inference on the Sphere

Abstract

Reconstructing a scene from images requires knowing where each photo was taken. Once camera orientations are fixed, the remaining step is translation averaging: estimating all camera positions given, for many camera pairs, the unit-length direction from one camera toward the other. These directions come from imperfect image matching, so a large fraction are not merely noisy but completely wrong (outlier). Existing solvers break down because none decides which measurements to trust: convex relaxations tolerate roughly a third of wrong directions, and robust local optimisers only refine their initial guess. Discrete optimisation is the classical remedy (give each variable a finite set of candidates and search globally for the combination explaining the most data, so the correct majority outvotes the wrong minority) but it needs a bounded candidate set, and camera positions live in unbounded 3D space. Our key idea is to change the question. A camera's position is unbounded, but the direction in which it is seen from a fixed reference point lies on the unit sphere, a bounded surface we cover with a fixed grid of cells. We plant reference points and, for each, globally assign to every camera the cell in which it appears from that point. Each assignment pins every camera to a ray, and robustly intersecting a camera's rays returns its 3D position. The proposed solver halves the median error of state-of-the-art solvers on synthetic graphs, and widens that margin to – at the outlier levels where convex relaxations break down. As a drop-in replacement for the global-positioning stage of the GLOMAP structure-from-motion pipeline, it improves or matches GLOMAP on every benchmark we test, wins end-to-end on ETH3D even after bundle adjustment, and, unlike its randomly initialised competitors, returns the same solution on every run. The code will be made public.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
translation averaging, global structure-from-motion

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_zZAPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses camera position recovery from noisy relative directions contaminated by incorrect matches, assuming known camera orientations. Its key idea is to introduce multiple virtual reference points, transforming an unbounded position search into finite direction labeling on the unit sphere. The method constructs Markov random fields and uses TRW-S to jointly infer camera directions, then recovers positions through robust ray triangulation and improves accuracy through continuous refinement. The reported experiments show that it outperforms several baselines on synthetic data with high outlier rates. When integrated into GLOMAP as a replacement for its positioning stage, it delivers substantial gains on ETH3D and smaller improvements on other datasets. Its main limitations are higher computational cost and continued sensitivity to weak geometric constraints.

Critical strengths

The paper represents each camera by its direction from multiple virtual reference points, replacing direct discretization of unbounded 3D positions with a shared, bounded spherical label space. Positions are then recovered through robust ray triangulation. The geometric characterization of compatible direction pairs enables sparse interactions within the discrete model. This is the paper’s strongest methodological contribution, and it provides a concrete way to apply discrete joint inference without requiring a predefined scene bounding box.

Critical weaknesses

In my view, the connection between the pairwise compatibility scores and the existence of a globally consistent camera configuration requires further clarification. After depths are eliminated independently for each edge, it is unclear to me how the model ensures that each camera has a consistent depth across all incident constraints. Certain direction assignments could therefore receive high pairwise scores without corresponding to a single, realizable configuration of camera positions. Whether exact per-edge elimination preserves equivalence to the original problem merits further discussion. Subsequent multi-seed fusion and continuous refinement may mitigate this issue, but I find the explanation of how they do so, and under what conditions, insufficient.

Questions for authors

For a given seed, could one constraint require a camera to lie closer to the seed along its assigned ray, while another requires it to lie farther away, such that both constraints are individually compatible but cannot be satisfied simultaneously? How does the algorithm detect and resolve such conflicts? After depths are eliminated independently for each edge, which step ensures that all direction assignments jointly correspond to a single, globally consistent configuration of camera positions? Does this require additional assumptions? The paper discusses outlier rates, geometric degeneracies, scale selection, and parameter sensitivity. However, I could not find a discussion of whether pairwise-compatible direction assignments are necessarily globally realizable. Could the authors clarify whether such conflicts can arise and, if so, how the method handles them?

1 comment

Thank you for this insightful question. After submission, we realized that the multiple seed points can be solved jointly in a single MRF run, thus reducing the runtime by 2x to 3x on average (up to 12.7x on Phototourism) as all K=4K = 4 non-coplanar seeds share one table per edge. Below, we first answer for our paper's per-seed solver (K=8K = 8), and then explain how the joint solver strengthens depth consistency.

1. Can a Single Seed Admit Conflicting Depths Across Incident Edges? Yes

Your geometric intuition is right. Fix one seed SkS_k and assign camera ii ray direction vk,i∈S2v_{k,i} \in S^2, so ci=Sk+rk,ivk,ic_i = S_k + r_{k,i} v_{k,i} (rk,i>0r_{k,i} \gt 0). Suppose camera ii has two neighbors j,mj, m at fixed 3D positions cj=Sk+rk,jvk,jc_j = S_k + r_{k,j} v_{k,j} and cm=Sk+rk,mvk,mc_m = S_k + r_{k,m} v_{k,m}:

  • On edge (i,j)(i, j), analytical depth elimination (Prop. 1, Cor. 1) checks only whether some ratio αij=rk,j/rk,i>0\alpha_{ij} = r_{k,j} / r_{k,i} \gt 0 aligns αijvk,j−vk,i\alpha_{ij} v_{k,j} - v_{k,i} with t~ij\tilde{t}_{ij}. Given rk,jr_{k,j}, exact alignment pins camera ii to depth rk,i(j)=rk,j/αij⋆r_{k,i}^{(j)} = r_{k,j} / \alpha_{ij}^\star.
  • On edge (i,m)(i, m), a second direction t~im\tilde{t}_{im} can be individually 100% compatible (αim⋆>0\alpha_{im}^\star \gt 0, θim=0\theta_{im} = 0) while requiring a different depth rk,i(m)=rk,m/αim⋆≠rk,i(j)r_{k,i}^{(m)} = r_{k,m} / \alpha_{im}^\star \neq r_{k,i}^{(j)} along the same ray vk,iv_{k,i}.

Thus, for one seed (K=1K = 1), per-edge depth elimination is an exact per-edge projection onto coplanarity and forward orientation that relaxes shared node depths across incident edges: sliding cic_i along vk,iv_{k,i} preserves span(vk,i,vk,j)\text{span}(v_{k,i}, v_{k,j}), leaving a 1-DoF radial ambiguity per camera and omitting cycle products ∏e∈Cαe=1\prod_{e \in C} \alpha_e = 1. In our paper's solver, per-seed MRFs do not enforce shared depths alone - consistency is restored across the KK seeds during multi-seed fusion. This is why the final robust triangulation is crucial.

2. Which Step Ensures Global 3D Realizability - and Under What Assumptions?

Multi-seed ray intersection (Sec. 3.5) across the $K $ seeds eliminates this radial ambiguity and ties every camera to one shared depth across all edges:

  • Locking in one shared 3D configuration: Requiring camera ii's rays across seeds Sk≠ciS_k \neq c_i to intersect at one 3D point ci∈R3c_i \in \mathbb{R}^3 uniquely pins every depth rk,i=∣ci−Sk∣r_{k,i} = \\|c_i - S_k\\| whenever ≥2\ge 2 seeds are not collinear with cic_i. Equivalently, on edge (i,j)(i, j), each seed SkS_k off (ci,cj)(c_i, c_j) confines cj−cic_j - c_i to a plane Πk,ij\Pi_{k,ij} containing t~ij\tilde{t}_{ij}. For any two seeds Sk,SlS_k, S_l not coplanar with (ci,cj)(c_i, c_j), Πk,ij\Pi_{k,ij} and Πl,ij\Pi_{l,ij} intersect in the unique line span(t~ij)\text{span}(\tilde{t}_{ij}). With α⋆>0\alpha^\star \gt 0, this enforces cj−ci=βijt~ijc_j - c_i = \beta_{ij} \tilde{t}_{ij} (βij>0\beta_{ij} \gt 0) and cycle closure ∑e∈Cβet~e=0\sum_{e \in C} \beta_e \tilde{t}_e = 0 for one shared 3D configuration (up to ρ=3.96∘\rho = 3.96^\circ).
  • Zero extra scene assumptions: Whenever the seeds span R3\mathbb{R}^3 (≥4\ge 4 non-coplanar seeds) no single plane contains all seeds, guaranteeing ≥2\ge 2 distinct planes per edge in R3\mathbb{R}^3.

3. How Conflicts Are Detected and Resolved: Original vs. Joint Solver

Conflicts are resolved in three steps:

  1. Consensus ray triangulation (Sec. 3.5): Because one seed's MRF can pick a depth-conflicting or outlier ray, for each camera ii we test candidate 3D centers from all K(K−1)/2=28K(K - 1) / 2 = 28 seed-ray pairs, count seeds agreeing within τ+ρ\tau + \rho, reject conflicting/outlier rays (breaking ties by lowest residual), and refine cic_i on inliers via Huber ray IRLS. This is the standard robust triangulation in both COLMAP and GLOMAP.
  2. Round-2 soft priors (Alg. 1): Reconciled Round-1 3D positions define soft priors ϕk,i\phi_{k,i} steering all K=8K = 8 seeds toward one shared depth-consistent 3D configuration in Round 2.
  3. Final continuous refinement: Inlier BATA refinement removes sub-cell (≤3.96∘\le 3.96^\circ) quantization error, yielding 0 failures across all 40 real scenes in Tables 1-3.

How the joint solver relates: Rather than reconciling rays only after independent TRW-S runs, our joint MRF couples a K=4K = 4 non-coplanar tetrahedron via cross-seed skew-ray potentials ψkl(vk,i,vl,i)\psi_{kl}(v_{k,i}, v_{l,i}) that penalize non-intersecting or negative-depth rays directly inside TRW-S, followed by the same triangulation and refinement (0 failures on 40/40 scenes, 34.9 ETH3D, 2x-3x faster). This is simple change to the algorithm but allows for directly considering multi-view consistency.

Reviewer_pt3vOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses translation averaging from noisy relative translation directions. This setting is a curial step in global structure from motion after the camera rotation averaging. In this paper each camera is represented by its direction from a few virtual reference points, producing a bounded spherical label space. For each reference point, a pairwise Markov random field encodes geometric compatibility between direction labels and is approximately optimized using TRW-S. The resulting directions are fused through robust triangulation, followed by continuous refinement.

The main contributions are the spherical labeling formulation, a sparse compatibility construction with analytical depth elimination, and an analysis of quantization error.

The authors validated the method with synthetic experiments together with integration into GLOMAP on three datasets.

Critical strengths

The paper proposed a substantive geometric formulation enabling bounded discrete inference. Representing camera positions through directions from virtual reference points provides a reusable spherical label space without requiring a scene bounding box. This is a meaningful modeling contribution that makes graph-wide discrete reasoning applicable to robust translation estimation.

Secondly, validation showing benefits within a complete SfM pipeline. The experiments evaluate both isolated positioning and reconstruction after bundle adjustment, alongside synthetic robustness tests. Particularly compelling is the improvement in ETH3D mean AUC@5° from 50.1 to 60.6 after bundle adjustment. This supports the practical importance of the proposed initialization: its benefits extend through subsequent reconstruction stages, strengthening the case beyond improvements on synthetic direction graphs.

Critical weaknesses

My first major concern is the connection between the discrete formulation and the original 3D positioning problem. The depth-elimination result establishes compatibility between two direction labels on an individual edge, but I do not see why satisfying these pairwise conditions necessarily implies a single, geometrically consistent set of camera positions across the graph. Also, TRW-S does not generally guarantee a globally optimal solution. I therefore remain unclear about what “global” means in this formulation and how it supports the paper’s explanation of improved robustness.

My second concern is how much of the improvement actually comes from discrete inference. The pipeline also includes continuous initialization, triangle-based confidence weights, multiple restarts, and continuous refinement. I would like to see a direct comparison that removes TRW-S while keeping the other components unchanged. Given that positioning takes roughly 10–100 times longer than the continuous baselines, it would also be useful to compare against multi-start continuous optimization under a similar computational budget. The current experiments do not sufficiently establish whether the additional complexity and cost of the discrete module are necessary.

Questions for authors
  1. Is the spherical MRF an exact reformulation of translation averaging, or a surrogate objective? In particular, how is consistency of the depth variables across different edges enforced or recovered? Please also clarify whether “global” refers to graph-wide approximate inference or a guarantee of global optimality.

  2. I would like to better understand how sensitive the method is to the number and placement of the reference points. The paper uses eight reference points with fixed directions and cyclic distances, while the seed-count ablation covers one dense synthetic setting. How reliable is this configuration for sparse view graphs or scenes whose extent is much larger than the gauge baseline? Could different reference-point layouts or distances substantially change the accuracy or failure rate? This would help me assess whether the chosen configuration is a reliable default across different reconstruction scenarios.

Reviewer_YHUCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses robust translation averaging under severe outlier contamination, where conventional continuous optimization can be sensitive to initialization and local minima. Its key idea is to reformulate the unbounded camera-position problem as discrete directional labeling on the sphere relative to multiple planted seeds. The method derives sparse pairwise compatibility constraints, performs per-seed MRF inference, and fuses the resulting rays through robust triangulation followed by continuous refinement. Experiments on synthetic and real-world datasets suggest improved robustness over existing translation-averaging approaches and effective integration into a global SfM pipeline.

Critical strengths
  1. The paper converts unbounded 3D translation averaging into bounded spherical labeling using multiple seeds, enabling joint discrete reasoning over camera directions while retaining a clear geometric interpretation.

  2. The compatibility-arc derivation, sparse pairwise construction, and multi-seed fusion form a coherent method, and experiments on both synthetic and real SfM benchmarks demonstrate robustness to substantial outlier contamination.

Critical weaknesses
  1. Each per-seed MRF has potentials that are not graph-cut representable and is solved with TRW-S, which provides only a lower bound and a weak-tree-agreement fixed point, but the title and paper describe the labeling as globally optimal. No duality gap or bound tightness is reported, and the seeds are fused only afterward, so it is unclear how often the returned labeling is optimal.

  2. The pipeline relies on a continuous pre-solution, multiple restarts, two outer rounds, IRLS refinement, robust triangulation, edge classification, and final BATA polishing, but the ablations in Appendix D.4 cover only the seed count and prior weight in a single setting. In addition, the baselines, GLOMAP-BATA, start from random initialization while the proposed method uses a continuous pre-solve. It is unclear how much of the gain comes from the MRF rather than from the surrounding stages.

Questions for authors
  1. Equation (8) minimizes an unnormalized cross-product residual, whereas the gate in Equation (9) uses the normalized angle, and the two minimizers differ under noise. How often does the gate reject a label pair for which another positive depth ratio satisfies the angular threshold, and how does using the angle-minimizing ratio change the results?

  2. Depth ratios are eliminated per edge, and cycle consistency (products equal to one) is not enforced. Is the spherical MRF an exact reformulation or a relaxation of the original problem?

  3. What fraction of the per-seed MRFs reach a vanishing TRW-S duality gap?

  4. Could the authors run the strongest continuous baseline from the same initialization and edge filtering, and report ablations of the outer rounds, IRLS refinement and final BATA polish?

2 comments

Thank you for the thorough review. After submission, we realized that the multiple seed points can be solved jointly in a single MRF run, thus reducing the runtime by 2x to 3x on average (up to 12.7x on Phototourism) as all K=4K = 4 non-coplanar seeds share one table per edge. Crucially for W1, this joint MRF enforces cross-seed 3D ray intersection during TRW-S rather than only afterward, lifting ETH3D AUC@5° to 34.9.

1. "Global" Wording, TRW-S Gaps (W1, Q3), and 3D Realizability (Q2)

You are right regarding optimality. To clarify, we never called our method "globally optimal" - we used "global" (and "optimised globally") to differentiate full-sphere discrete search from iterative methods starting from an initialization (e.g., IRLS). To avoid any ambiguity, we will remove "global" from the title and elsewhere where it may be misleading.

  • TRW-S LP duality gaps (W1, Q3): Eq. (9) is non-submodular, so TRW-S maximizes Kolmogorov's (2006) LP bound DD without guaranteeing zero gap (E−D)/E(E - D) / E on loopy graphs (verified <10−15\lt 10^{-15} on trees). Across all 320 per-seed MRFs on all 40 scenes (8 seeds ×\times 40 scenes, 50 sweeps):
LP Gap MetricHaiper (24)ETH3D (104)Herit. (24)EuRoC (48)Photo. (120)All (320)
Gap below 1e-416.7%12.5%0.0%0.0%0.0%5.3%
Gap below 1e-216.7%18.3%8.3%0.0%0.0%7.8%
Median (E - D) / E3.4%8.9%10.6%40.7%45.2%27.2%

We will report LP gap distributions in the supp. mat.

  • Single-seed relaxation vs. multi-seed realizability (Q2): Per edge (i,j)(i, j), depth elimination is exact: (va,vb)(v_a, v_b) admit ri,rj>0r_i, r_j \gt 0 aligning cj−cic_j - c_i with t~ij\tilde{t}_{ij} iff coplanar with t~ij\tilde{t}_{ij} and forward (α⋆>0\alpha^\star \gt 0). Over loopy graphs, one seed's MRF is a relaxation: depth rk,ir_{k,i} is unconstrained across incident edges, omitting cycle products ∏e∈Cαe=1\prod_{e \in C} \alpha_e = 1. Intersecting rays across K≥4K \ge 4 non-coplanar seeds restores 3D realizability and cycle closure. Each seed SkS_k off (ci,cj)(c_i, c_j) confines cj−cic_j - c_i to a plane containing t~ij\tilde{t}_{ij}. For any two seeds not coplanar with (ci,cj)(c_i, c_j), their two planes intersect in the line along t~ij\tilde{t}_{ij}, pinning every camera to one shared center cic_i with ∑e∈Cβet~e=0\sum_{e \in C} \beta_e \tilde{t}_e = 0 up to ρ=3.96∘\rho = 3.96^\circ.

2. Cross-Product Ratio Eq. (8) vs. Angle-Minimizing Ratio in Eq. (9) (Q1)

Across all 4.12×1074.12 \times 10^7 label pairs on our 642-vertex icosphere (ρ=3.96∘\rho = 3.96^\circ), we compared Eq. (8)'s α⋆\alpha^\star to the angle minimizer αang=arg⁡min⁡α∈(0,∞)θ(α)\alpha_{\text{ang}} = \arg\min_{\alpha \in (0, \infty)} \theta(\alpha) inside the compatibility tube and baseline gate (∣pij∣≥0.1\\|p_{ij}\\| \ge 0.1):

  1. How often does the gate reject a valid pair? At cutoff τ+2ρ=12.93∘\tau + 2\rho = 12.93^\circ (τ=5∘\tau = 5^\circ), αang\alpha_{\text{ang}} accepts 5.59% of all pairs vs. 5.08% for α⋆\alpha^\star, rejecting 0.51% of all pairs (9.46% of accepted pairs, and 0.06% / 4.15% at raw 5°). Rejected pairs sit only at the θ≈τ\theta \approx \tau tail where Gaussian weight e−θ2/(2σ2)≈4×10−6e^{-\theta^2 / (2\sigma^2)} \approx 4 \times 10^{-6} is negligible.
  2. How does using αang\alpha_{\text{ang}} change the results? On accepted pairs, median θ(α⋆)−θ(αang)\theta(\alpha^\star) - \theta(\alpha_{\text{ang}}) is 0.014° (0.001° at 5°) - 280x below ρ=3.96∘\rho = 3.96^\circ - leaving TRW-S labelings and pose accuracy unchanged. Conversely, unrestricted inf⁡α>0θ(α)\inf_{\alpha \gt 0} \theta(\alpha) admits boundary limits (α→0+\alpha \to 0^+ or ∞\infty) on 26.0% of pairs that collapse cameras onto SkS_k, which Eq. (8) prevents.

3. Matched-Init Baseline and ETH3D Stage Ablations (W2, Q4)

On ETH3D (13 scenes, isolated positioner), removing one stage at a time from our full solver shows spherical MRF drives the gain (+7.6 AUC@5°) while every stage improves the final result:

ETH3D Ablation (13 scenes)AUC@5°AUC@10°vs. Full
Default GLOMAP (random init, 7s)27.235.8-7.7 / -6.8
Full minus Spherical MRF (same pre-solve + filter)27.335.8-7.6 / -6.8
Full minus Per-seed ray IRLS (TRW-S + triang. only)34.442.0-0.5 / -0.6
Full minus 2nd outer round (1 round, 16s)34.642.4-0.3 / -0.2
Full minus Continuous pre-solve (uniform prior)34.742.1-0.2 / -0.5
Full minus Final BATA polish34.741.7-0.2 / -0.9
Ours (joint, full, 42s)34.942.6best

Starting Default GLOMAP from our exact pre-solve and filter without the MRF adds only +0.1 AUC@5° (27.3), whereas our MRF reaches 34.7 even without the pre-solve and 34.9 / 42.6 with all stages active.

Thank authors for the detailed response and the new experiments. I read the replies to all four reviews.

  1. “Global” and TRW-S (W1/Q2/Q3):

    Thank you for removing “global” and the LP gap report. The gap is small for only 5.3% of the per-seed MRFs (median 27.2%, >40% on EuRoC and Phototourism), so I still cannot tell how close the returned labelings are to optimal.

  2. Eq. (8) vs (9) (Q1): The reply is clear. Please add the gate statistics to the appendix.

  3. Ablation (W2/Q4):

    3.1. The ETH3D table addresses my main concern. Starting default GLOMAP from the same pre-solve and filter gives 27.3 vs 27.2, whereas the full method gives 34.9, so the gain comes from the spherical MRF and not from the surrounding stages. Please add this table to the paper.

    3.2. It covers only ETH3D and the joint variant introduced after submission. The gains on the other benchmarks are small (29.8 vs 29.8 on Heritage, 32.6 vs 32.9 on Haiper), so the same ablation on one more dataset would help.

Reviewer_q5THOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a discrete optimization framework for translation averaging by representing camera positions as spherical direction labels relative to multiple reference seeds. The method uses pairwise geometric compatibility, MRF inference with TRW-S, and robust multi-view triangulation to recover camera positions.

Critical strengths

1 .The spherical direction labeling formulation is an interesting approach to transforming an unbounded camera position estimation problem into a bounded discrete optimization problem.

  1. The paper provides geometric analysis of pairwise compatibility, depth elimination, and quantization error, supporting the proposed formulation.
Critical weaknesses
  1. Most baselines are classical translation averaging methods from 2014–2018 or variants of BATA. Recent relevant methods, particularly TriP [1], Cycle-Sync [2], and Adaptive Annealing [3], are cited or discussed but not included in the experiments. Notably, TriP and Cycle-Sync have reported strong localization accuracy as well as computationally efficient results on ETH3D, which is also the benchmark where the proposed method demonstrates its largest gains. Without comparisons under consistent evaluation settings, it is difficult to establish the method's competitiveness against recent approaches.

  2. The proposed discrete solver requires one to two orders of magnitude more computation than the evaluated continuous baselines. However, the paper does not sufficiently analyze whether the reported accuracy improvements justify this additional computational cost, particularly given the limited gains after bundle adjustment on some benchmarks. A more detailed accuracy–runtime analysis would help establish the practical value of the proposed approach.

  3. The related work discussion could be expanded to include additional relevant studies, such as TriDE [4] and Fusing Directions and Displacements in Translation Averaging [5]. These works address related aspects of translation-direction consistency and translation averaging. Discussing their relationship to the proposed approach would help better contextualize its contributions.

References

[1] Fan, Z., Li, W., Wang, J., and Shi, Y. TriP: A Triangle Puzzle Approach to Robust Translation Averaging. arXiv:2605.07143, 2026. https://arxiv.org/abs/2605.07143

[2] Li, S., Shi, Y., and Lerman, G. Cycle-Sync: Robust Global Camera Pose Estimation through Enhanced Cycle-Consistent Synchronization. NeurIPS, 2025. https://arxiv.org/abs/2511.02329

[3] Chitturi, S., and Govindu, V. M. Adaptive Annealing for Robust Averaging. ECCV, 2024. https://www.ecva.net/papers/eccv_2024/papers_ECCV/html/8834_ECCV_2024_paper.php

[4] Chen, F., Wang, Y., and Shi, Y. TriDE: Triangle-Consistent Translation Directions for Global Camera Pose Estimation. arXiv:2605.06889, 2026. https://arxiv.org/abs/2605.06889

[5] Manam, L., and Govindu, V. M. Fusing Directions and Displacements in Translation Averaging. 3DV, 2024. https://ee.iisc.ac.in/cvlab/research/fusedta/

Questions for authors
  1. Could the authors include comparisons with recent robust translation averaging methods, particularly TriP, Adaptive Annealing, and the location estimation component of Cycle-Sync, in terms of both accuracy and runtime? Evaluations on ETH3D under consistent inputs, metrics, and computational settings would be particularly informative.
  2. In Eq. (2), fixing the gauge also enforces the selected seed edge's measured translation direction as exact. Since the relative directions are noisy, could the authors clarify how this hard constraint affects the recovered camera positions, particularly when the selected seed edge has a small angular error?
  3. Could the authors provide an ablation that isolates the contribution of spherical discrete inference from the other components, such as triplet-based confidence weighting, gauge-edge restarts, and the final BATA refinement? This would help clarify which component contributes most to the reported robustness improvements.
4 comments

We thank for the thorough review. After submission, we realized that the multiple seed points can be solved jointly in a single MRF run, thus reducing the runtime by 2x to 3x on average and up to 12.7x on large scenes (Sacre Coeur: 2351s to 186s). Coupling multiple non-coplanar seeds in one MRF shares one compatibility table per edge and enforces 3D ray intersection inside TRW-S, keeping accuracy similar while accelerating the solve.

1. Comparisons with TriP, Cycle-Sync, and Adaptive Annealing

Crucially, TriP [1], Cycle-Sync [2], and Adaptive Annealing (AA) [3] never run on the nowadays-standard joint point+camera SfM graphs in their papers. All three test solely on small camera-only graphs, whereas modern global SfM (GLOMAP) positions cameras and 50k-130k 3D tracks jointly. With no public code for TriP or AA, we implemented both in C++ and ran official Cycle-Sync across all 40 real scenes:

  • Cycle-Sync: Official code builds dense O(n2)O(n^2) arrays and solves (3n+3)×(3n+3)(3n + 3) \times (3n + 3) KKT systems (O(n3)O(n^3)). Under a 10-min/scene cap on point+camera graphs (nn up to 8,038 ETH3D [4.7 GB], 62,536 Heritage/EuRoC [281 GB], 129,530 Photo. [1.2 TB]), it OOMs (20) or exceeds 10 min (16) on 36/40 scenes (only 4 scenes with n<2,000n \lt 2{,}000 finish, taking 5.5-9 min on ETH3D). Even on camera-only Phototourism (N≤911N \le 911), it averages 80.5s (up to 173s). Thus, running Cycle-Sync in the standard point+camera setting is impossible on modedate-to-large scenes.
  • TriP: Chaining rigid 3-cycles fails on 29/40 point+camera scenes (0/15 Phototourism, 0/6 EuRoC) since camera-to-point edges form zero triangles (3D points have no mutual edges).
  • Camera-only + BA and AA: Dropping 3D tracks (camera-only) lets [1, 2, 3] finish in 1-6s, but loses over 25 AUC@5° points after BA on ETH3D. On joint graphs, single-basin AA trails our solver on every dataset.

Positioning only (point+camera, AUC@5° and mean time):

MethodETH3D (13)Haiper (3)Heritage (3)Photo. (15)EuRoC (6)
Default GLOMAP27.2 (7s)32.6 (4s)29.8 (61s)17.3 (266s)44.7 (188s)
Adaptive Annealing [3]22.1 (4s)29.9 (1s)9.6 (6s)9.4 (26s)24.3 (26s)
Cycle-Sync [2]2.8 (2/13 ok)7.6 (2/3 ok)-- (0/3 ok)-- (0/15 ok)-- (0/6 ok)
TriP [1]23.5 (7/13 ok)31.5 (8s)6.3 (1/3 ok)-- (0/15 ok)-- (0/6 ok)
Ours (original)34.5 (54s)33.0 (13s)29.9 (156s)18.8 (1066s)46.0 (678s)
Ours (joint)34.9 (42s)32.9 (5s)29.8 (125s)18.8 (379s)46.1 (463s)

2. Related Work: TriDE and Manam & Govindu

TriDE [4] denoises edge directions on 3-cycles before positioning, whereas we infer camera directions from virtual seeds without clean triangles. Manam and Govindu [5] fuse direction and displacement costs in continuous IRLS, complementing our discrete init. We include both in Sec. 2.

3. Gauge-Edge Noise in Eq. (2) and ETH3D Ablation

Rotations RiR_i fix world orientation, 4 high-triplet restarts keep seed-edge error below 3.96°, and final continuous refinement unpins cj⋆c_{j^\star} (fixing only ci⋆=0c_{i^\star} = 0 and scale) so cj⋆c_{j^\star} relaxes freely. Removing one step at a time on ETH3D confirms that spherical MRF drives the gain (+7.6 AUC@5°) and every step improves accuracy:

ETH3D Ablation (13 scenes)AUC@5°AUC@10°vs. Full
Default GLOMAP (baseline, 7s)27.235.8-7.7 / -6.8
Full minus Spherical MRF (same init + filter)27.335.8-7.6 / -6.8
Full minus Triplet weights34.541.9-0.4 / -0.7
Full minus Multi-edge restarts (1 restart, 10s)34.542.0-0.4 / -0.6
Full minus Continuous pre-solve34.742.1-0.2 / -0.5
Full minus Final BATA refinement34.741.7-0.2 / -0.9
Ours (joint, full, 42s)34.942.6best

Thank you for the detailed rebuttal and additional experiments. Most of my concerns have been addressed, and I have one final question regarding the comparisons with TriP and Cycle-Sync.

Since both methods were originally designed for camera-only graphs, could the authors evaluate them under their intended settings on the same benchmark? The current results on joint point-camera graphs may not reflect their actual performance.

If the proposed method still demonstrates a clear advantage under a fair comparison, I would be willing to raise my score to 3.

Thank you for engaging! Below are the updated tables across all five benchmarks, evaluating TriP [1], Cycle-Sync [2], and AA [3] in their camera-only setting.

MethodETH3D (13)Haiper (3)Heritage (3)Photo. (15)EuRoC (6)
Default GLOMAP27.2 (7s)32.6 (4s)29.8 (61s)17.3 (266s)44.7 (188s)
Adaptive Annealing [3]22.1 (4s)29.9 (1s)9.6 (6s)9.4 (26s)24.3 (26s)
Adaptive Annealing [3] (camera-only)22.4 (0.02s)26.1 (0.01s)12.1 (0.04s)16.7 (41s)11.5 (5s)
Cycle-Sync [2]2.8 (2/13 ok)7.6 (2/3 ok)-- (0/3 ok)-- (0/15 ok)-- (0/6 ok)
Cycle-Sync [2] (camera-only)25.4 (1.6s)29.0 (0.4s)20.1 (2.9s)17.1 (81s)22.9 (11s)
TriP [1]23.5 (7/13 ok)31.5 (8s)6.3 (1/3 ok)-- (0/15 ok)-- (0/6 ok)
TriP [1] (camera-only)31.6 (0.1s)26.8 (0.01s)13.1 (0.2s)17.6 (20s)11.5 (0.1s)
Ours (original)34.5 (54s)33.0 (13s)29.9 (156s)18.8 (1066s)46.0 (678s)
Ours (joint)34.9 (42s)32.9 (5s)29.8 (125s)18.8 (379s)46.1 (463s)

On camera-only graphs, TriP, Cycle-Sync, and AA finish on all 40/40 scenes. On dense unordered collections (ETH3D, Phototourism), Cycle-Sync and TriP works decently but still lags behind point+camera default GLOMAP. The biggest difference is on EuRoC which is sequential trajectory where the camera-only constraints are degenerate for scale estimation.

Thank you for the additional experiments and clarifications. All my concerns have been addressed.

Comments on the paper

1 comment

We thank the four reviewers. Two notes concern all of them which we answer here. We also answer each reviewer separately.

  1. The word "global" (YHUC W1, pt3v W1). By "global" we mean that the search covers the whole label space, as opposed to a local descent from a random start. We do not mean globally optimal, and the paper claims no optimality guarantee. To remove the ambiguity the title is now Translation Averaging as Discrete Inference on the Sphere, and the phrase "optimised globally by TRW-S" is reworded where it occurs.
  2. A joint variant. After submission we found that the seeds can be solved in one MRF: a cross-seed term makes a camera's rays meet during inference. Accuracy is unchanged and positioning runs significantly faster.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.