-
The usefulness of the method is not entirely supported, due to many factors. While the obtained reconstructions are indeed slightly better, let's recall that a representation's quality is usually measured by how useful it is for solving downstream tasks; most representation learning methods are usually benchmarked on metrics such as classification accuracy or generation quality. However, the paper only evaluates its learned representations in downstream tasks in the generation experiment of Section 5.2, which presents several weaknesses. Firstly, note how the reconstructions in Table 1 compare e.g. a 3DS2VS baseline trained by the authors, against the CanonAE-version of 3DS2VS also trained by the authors, which is correct. For generation, however, the authors pick an off-the-shelf checkpoint of 3DS2VS for their baseline computations, compared against their CanonAE version of 3DS2VS. This in itself is a problem because we can't attribute gains to either of the architectures since they were trained in potentially different settings (different learning rates, training iterations, batch sizes, optimization differences, etc). This weakens the comparison in Section 5.2. Not only that, but the generation quality is assessed only qualitatively. We don't have a quantitative metric comparing CanonAE against the 3DS2VS baseline. Finally, the generation is studied (for some reason) on ShapeNet, which is not among the three datasets presented as the scope of the paper's empirical evaluation. Why is there no corresponding ShapeNet reconstruction comparison? The paper also gives no experimental details or quantitative results on how the ShapeNet CanonAE poses and latents were obtained. Overall, this weakens the only downstream-task evidence that the paper provides. Furthermore, such evidence covers a single dataset (ShapeNet), a narrow-scoped evaluation.
-
Table 1 and ablation results report no error bars or standard deviations over multiple seeds. Many differences are very small, and therefore they might be within noise. It is therefore unclear whether the improvements are real, and whether some components of the method are doing real work. In this regard, the ablation on 5.4.2 on no alternate optimization shows that this component barely bears any difference in reconstruction. A more simple version of this method would be that of using pose synchronization as a preprocessing step, and then running any kind of AE normally. This is arguably simpler and more elegant, and provides most if not all the benefits.
-
The main body of the paper lacks mathematical clarity. Several formulas are introduced without prior definitions of their components. For instance, equations (1) and (2) appear in the introduction with many undefined terms, and Sections 3 and 4 (which represent the in-depth method sections) never reintroduce them formally. The confusion carries through the paper; for instance, Section 4 opens by proposing a “scalable approach” to solving equation (2), but the reader never had a chance to grasp equation (2), since it was never properly derived or defined. A critique in the same spirit is that the problem setting is never made explicit, which makes every section hard to follow. For instance, Section 3.1 is the first method section, where the setting should be established. It starts by assuming a fixed connected weighted graph, with no explanation of whether the graph represents e.g. a datapoint or some other encoder property of the data that the authors aim to study. In this same section, losses and the “stationary alignment” are also defined informally, and it is unclear why they are introduced or how they connect to the exact problem that the authors aim to solve, since the problem itself has not yet been defined clearly and formally. The reader can probably infer the method despite these obstacles, but overall the writing and clarity are of very poor quality.
-
The paper shows how pose synchronization can plausibly aid learning in vanilla autoencoders. On the other hand, I believe group-invariant/equivariant autoencoders attack the same problem through architectural invariance rather than explicit pose alignment, e.g. IE-AE [1], SGM [2], and RECON [3] (incidentally, all these works are missing from the related work section and are relevant to both the “Joint shape-and-pose learning” and “Canonicalization of 3D collections” subsections). Since these methods use a group-invariant encoder, their reconstruction loss is be independent of input orientation (I believe this to be true?), in which case CanonAE would provide no benefit. A comparison is needed, at a minimum theoretically and ideally empirically: does CanonAE bring any advantage over such methods, and how large is the practical gap in representation learning performance? This also relates to L123: “These methods jointly model intrinsic shape variation and pose, making optimization susceptible to pose ambiguity and poor local minima.” This may be true in theory. But does it translate into empirical disadvantages relative to pose synchronization? In practice, such methods might be competitive with approaches that build in group equivariance/invariance priors. Showing this is important to convince the reader of the method’s value.
-
Appendix D.1 introduces several critical assumptions whose failure modes are not discussed. Where can they fail? For instance, are there specific instances in any of the experiments where they break, and how does that affect the behavior of the method?
-
No analysis of the number of neighbors k is provided. It is mentioned that k is chosen large enough to satisfy the convergence conditions of App. H, but it is unclear how to choose k in arbitrary scenarios. A sensitivity analysis on k would enable practitioners to apply the method succesfully. So would a discussion of what happens, theoretically and empirically, when a “wrong” k violates the convergence assumptions.
-
Section 5.1 mentions that, at test time, each shape is aligned to its 16 nearest training shapes, unlike the single forward pass of other canonicalization/AE methods. This test-time cost should be reported as part of the price paid, e.g., throughput of this method vs. the baselines.
References:
[1] Winter, R., Bertolini, M., Le, T., Noe, F., & Clevert, D. A. (2022). Unsupervised learning of group invariant and equivariant representations. Advances in Neural Information Processing Systems, 35, 31942-31956.
[2] Allingham, J. U., Mlodozeniec, B. K., Padhy, S., Antorán, J., Krueger, D., Turner, R. E., ... & Hernández-Lobato, J. M. (2024). A generative model of symmetry transformations. Advances in Neural Information Processing Systems, 37, 91091-91130.
[3] Urbano, A., Romero, D. W., Zimmer, M., & Pokutta, S. (2026, April). RECON: Robust symmetry discovery via explicit canonical orientation normalization. In International Conference on Learning Representations (Vol. 2026, pp. 7131-7163).