- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 3
Read review
Paper summary
The paper proposes a new method to maximize consensus rotation averaging. Their method uses a discretization of SO(3). The search over this discretization becomes a pairwise Markov Random Field. They then use shrinking upper and lower bounds to converge to a solution. In theory, they get exact recovery under some conditions on the inlier and outlier sets of pairwise measurements. Experiments demonstrate that their method outperforms some existing methods in some regimes.
Critical strengths
The method outperforms many existing algorithms for robust rotation synchronization. These include other message-passing algorithms and energy minimization methods on a variety of datasets.
The direct optimization of a consensus measure seems to have been pursued only in the context of random-sampling methods (such as RANSAC). The use of the MRF appears novel and is an interesting approach to getting robust rotation synchronization to work.
Critical weaknesses
In its current state, the paper is quite difficult, if not impossible, for a human to read. Jargon is frequently used and undefined, the text does not clarify as it goes in a way that is natural for a human reader, and there is not enough signposting/roadmaps to really get a sense of what is going on. Furthermore, I am left wondering what has really been proven and why it matters.
The method considered is a rotation averaging method. For large-scale 3D reconstruction, methods have moved towards foundation model-based approaches. The authors do not sufficiently motivate why we should still care about these incremental solvers in this larger context. In fact, on all datasets tested, state-of-the-art methods are typically found by these learned methods, and they are also quite robust. This is especially relevant to a conference like ICLR, and I therefore do not find it to be a good and modern fit.
Questions for authors
Could the authors expand on their review of related work? It appears to be mostly AI-generated, and many references have been missed historically in the development of robust rotation synchronization.
Also, it would be nice to have a better idea of the relevance of these ideas to the modern structure-from-motion and 3d vision landscape, and why these developments are important and useful.
Thank you for your thorough review.
1. Global vs. Regional Certificates and Fixing Section 3.4
You are right: when Stage B prunes the 30° grid around the Stage A estimate instead of keeping all of SO(3)N, closing the bound gap certifies a regional optimum over the retained region U, not a global one. We will retitle the paper "Certified Consensus for Robust Rotation Averaging", revise the abstract and Sec. 3.4 to separate global (U=SO(3)N) from regional (U⊊SO(3)N) certificates, and report the exact counts:
Fixing Section 3.4's fallback rule. Relaxing pruning only when branch-and-bound leaves an open gap cannot detect when pruning dropped a camera's true cell outside U and search converged inside U (9/20 runs at 50% outliers and 12/20 at 70% in App. B.2). We remove that claim and add a greedy coordinate sweep (Iterated Conditional Modes, ICM, 30 restarts) between Stages A and B that re-seats every camera over all 30° root cells without pairwise tables (+1.4–2.6 s). Stage B then builds the region around the swept incumbent. On the 20 sparse graphs at 50% outliers the full pipeline still closes 20/20 (median 60 s, max 101 s), and the certified count now matches or exceeds the ground-truth count on 20/20 instead of 11/20 (0.56° error). At 70% outliers it also closes 20/20 (median 76 s, max 111 s), with the certified count below ground truth on 2/20 instead of 12/20 (0.74° error). App. B.2 will report the 2 remaining cases (1 and 4 edges short).
2. Estimation vs. Certification Runtime and Hardware Parity
Table 1 timed Stage A (
GPU + CPU), while Table 5 ran the full loop onCPU(64.4 s mean / 65.2 s med. at 50%: 9.4 s Stage A + 55.0 s Stages B-C, and 74.1 s / 72.2 s at 70%). RunningGPU + CPUcloses 20/20 regional certificates at N=100 in 1.28 s (0.38 s Stage A + 0.90 s Stages B–C, 1.98 s with ICM). Mean runtimes on sparse 50%-outlier graphs (20 seeds, i9-10900K CPU, RTX 4090 GPU, †error above 5°):At N=1000 (183 s = 102 s Stage A + 81 s Stages B-C), 17/20 close to 0 gap and 3/20 stop at level 16 with a 1-edge (0.01%) gap - still 2–5x faster than GNC. Sec. 4.1 will state this.
3. When IRLS Polishing Leaves the Search Region (Algorithm 1)
Correct: IRLS polishing (Line 9) is unconstrained. In global mode (U=SO(3)N), Rpol∈U holds automatically. In regional mode (U⊊SO(3)N), Rpol can leave U and achieve an inlier count Cinc above the upper bound U of U, missing Line 11's equality test U=Cinc. Because U (Line 8) bounds every rotation in U before polishing, changing Line 11 to Cinc≥U proves R^ beats or ties all of U - certifying optimality over U∪{R^} without constraining IRLS, with max(U,Cinc) as the anytime upper bound. Our C++ solver already checks this bound and, if search continues, inserts the next-level cell containing R^i before rebuilding tables (we will update Alg. 1 to match).