Pith. sign in

REVIEW 4 major objections 6 minor 42 references

Decomposing affine transforms into bounded, probabilistically regressed components makes spatial transformers robust on fine-grained moth classification.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 16:53 UTC pith:T75FOIOJ

load-bearing objection Promising STN extension for transformers, but a core constraint formula is inverted and the code isn't available—needs review, not rejection. the 4 major comments →

arxiv 2509.11218 v1 pith:T75FOIOJ submitted 2025-09-14 cs.CV cs.AI

Geometrically Constrained and Token-Based Probabilistic Spatial Transformers

Johann Schmidt , Sebastian Stober This is my paper
classification cs.CV cs.AI
keywords spatial transformer networksfine-grained visual classificationcanonicalizationaffine transformationsprobabilistic regressiongeometric robustnessmoth classificationdata augmentation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Fine-grained visual classification remains brittle under rotation, scale, and perspective changes. The paper argues that Spatial Transformer Networks, often seen as fragile, can serve as effective canonicalizers in transformer pipelines if the affine transformation is decomposed into interpretable components—rotation, scaling, shearing—each regressed under explicit bounds and with a Gaussian variational posterior. This enables sampling-based canonicalization at inference and a component-wise alignment loss that exploits ground-truth augmentation parameters. On two fine-grained moth benchmarks, the method consistently outperforms both standard STNs and augmented training, improving robustness on heavily transformed test sets. The paper positions this as a flexible, backbone-agnostic alternative to equivariant architectures and data augmentation.

Core claim

The central claim is that predicting affine parameters component-wise—rotation angle, anisotropic scale, and shear—with bounded output domains and a Gaussian posterior over each component yields a substantially more robust spatial transformer than predicting the full matrix directly. The localization encoder is a shallow transformer that consumes the vision tokenizer's output, and the predicted components are composed into a matrix whose inverse is used to canonicalize the image. A component-wise alignment loss, computed against the known augmentation parameters, guides the encoder to the correct components. The paper reports top-1 accuracies of 97.6% on the EU-Moth test set and 73.4% on the

What carries the argument

The central object is the decomposed affine matrix M = R(θ)·S(sx,sy)·H(hx,hy), with each component regressed by a separate head under explicit bounds. Each component is treated as a Gaussian random variable; sampling composes a stochastic canonicalization. A component-wise alignment loss against ground-truth augmentation parameters provides geometric supervision. The localization encoder is a token-based transformer sharing the frozen tokenizer with the classifier, avoiding redundant pixel-space encoding.

Load-bearing premise

The component-wise alignment loss assumes the training augmentation applies rotation, scaling, and shearing in exactly the order R(θ)·S(s)·H(h) used in the decomposition; if the augmentation order differs, the regressed components have mis-specified targets and the loss no longer guides the localization encoder correctly.

What would settle it

If one re-runs the reported moth experiments with a different augmentation order (e.g., shear applied before scale) and sees no drop in robustness, then the component-wise alignment loss is not actually exploiting the assumed order; alternatively, if a deterministic, non-probabilistic version of the same component-wise regressor matches or beats the reported results, the posterior sampling is not the load-bearing ingredient.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the reported results hold, STNs become a viable, backbone-agnostic way to add geometric robustness to transformer classifiers without architectural changes.
  • The approach can be ported to other fine-grained or recognition tasks where objects appear under rotation, scale, and perspective variation.
  • The component-wise alignment loss requires augmentation parameters, meaning the method currently applies only to datasets or pipelines that provide such supervision.
  • Sampling eight transformations at inference gives a good accuracy–robustness trade-off; more or fewer samples reduce accuracy.
  • The Gaussian posterior variant outperforms the heavier-tailed Gamma-hyperprior variant on most benchmarks, suggesting a simple Gaussian suffices for the considered data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the gains replicate on other fine-grained datasets, the combination of explicit geometric decomposition and mild posterior sampling may close much of the gap to equivariant networks at a fraction of their cost.
  • Because the paper disables the KL term and still achieves its best numbers, the variational framing may be incidental; a deterministic constrained component-wise regressor could carry most of the benefit—an extension the paper does not test.
  • The alignment loss depends on knowing the augmentation order; the paper does not specify it, so the method's practical applicability outside carefully controlled augmentation pipelines is uncertain.
  • The approach could be extended to translation and reflection components, which the paper leaves to future work; if those components behave as well as rotation, scale, and shear, the method would cover the full affine group.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a spatial-transformer-based canonicalization module for transformer-based fine-grained visual classification. The method decomposes planar affine transformations into rotation, anisotropic scaling, and shearing; regresses each component under explicitly bounded domains using a shared frozen tokenizer and a 2-layer localization encoder; and adds Gaussian stochasticity by sampling transformation parameters during training and inference. A component-wise alignment loss uses augmentation parameters to supervise the regressed components. Experiments on EU-Moth and ECU-Moth, including roto-scaled and sheared test sets, compare the method against vanilla STN, Head-STN, Polar/LogPolar, Diffeomorphic STN, P-STN variants, and augmented training. The paper reports consistent robustness gains, with the best configuration being the Gaussian version with S=8 Monte Carlo samples.

Significance. If the reported results hold, the paper makes a useful empirical contribution: it demonstrates that a lightweight, backbone-agnostic STN-style canonicalizer can improve the geometric robustness of a pretrained Swin-Base classifier, and it provides a component-wise ablation that isolates the value of decomposed regression, probabilistic sampling, and a token-based localization network. The explicit limitations section is candid about the method's dependence on augmented training and ground-truth transformation parameters. However, confidence in the central claim is currently limited by three load-bearing issues: the component constraints in Eq. (7) are mathematically inconsistent as written, the KL term presented in the objective is disabled during actual training, and Table 1 contains no error bars or repeated-seed statistics. These issues must be resolved before the empirical conclusions can be fully credited.

major comments (4)
  1. [§4, Eqs. (13)-(14); §5.1] The component bounds in Eq. (7) are internally inconsistent as written. With s = 1 - (1/λs)sigmoid(·), the range is [1 - 1/λs, 1]. The text states λs ∈ (0,1) 'to ensure the strict positiveness of s', but positivity requires λs > 1; with the stated λs = 0.25 the scale head can emit values in [-3,1], including negative scales and hence reflections, directly contradicting the paper's exclusion of reflections. Similarly, h = (π/λh)tanh(·) with λh = 0.25 gives a shear range of [-4π,4π], not a 'reasonable' bound. This looks like a reciprocal typo (likely h = λh π tanh and λs ≈ 2 if only zoom-out scales are intended), but since no code link is provided, a reader cannot verify which formula was actually implemented. Because the constrained domains are load-bearing for the 'geometrically constrained' claim, Eq. (7) must be corrected and the actual implementation described.
  2. [§4, Eqs. (13)-(14); §5.1] The variational objective in Eq. (13) includes the KL divergence, but §5.1 states that the KL term was disabled entirely because it diminished test accuracy. The final training loss is therefore L_NLL + λ L_align, not Eq. (13). This is not merely a cosmetic discrepancy: without the KL term, pφ(T|x) is not a standard variational posterior, and the unit-Gaussian priors q(t) in Eq. (14) play no role in training. The paper should either present the actual objective and reframe the method as stochastic canonicalization, or justify why the KL term can be omitted while retaining the variational interpretation.
  3. [Table 1] Table 1 reports single top-1 accuracy values per condition with no error bars, repeated seeds, or significance tests. The central claim that the method 'outperformed all other baselines on all six benchmarks' rests on small differences (e.g., EU† 96.3 vs 95.4/95.2; ECU 73.4 vs 72.9/72.8), which may be within run-to-run noise on datasets of only 1445-1650 samples with 675/200 classes. Please report mean ± std over at least three seeds, or otherwise justify the stability of the reported numbers. This is important because robustness improvement is the paper's main empirical claim.
  4. [§5.2; Eq. (3); Eq. (12)] The alignment loss L_align compares each predicted component (θ, s, h) to augmentation parameters, but the paper never specifies the order in which the spatially augmented training set applies shearing, scaling, and rotation. Equation (3) fixes the decomposition M = R(θ)S(s)H(h); if the augmentation composes operations in a different order, the individual targets are not the components of Eq. (3), so the loss mis-specifies the regression targets. The test-set construction mentions shearing before the orbit-space construction, but the training augmentation order is not stated. Please specify the exact composition order used to generate augmentation labels and confirm that it matches Eq. (3).
minor comments (6)
  1. [§5] The text says 'complete implementation details publically available in our source code', but no repository URL or code link is provided. Please include it for reproducibility.
  2. [§4, Eqs. (7) and (12)] The symbols λθ, λs, λh are overloaded: in Eq. (7) they are co-domain bounds, while in Eq. (12) they are alignment-loss weights with different values. Use distinct notation to avoid confusion.
  3. [§4, Eq. (15)] Equation (15) appears to be a typo: p_{c,φ}(y|x) ≈ ε p_c(y|φ(x)x) should involve φ(x)^{-1}x, and the residual error ε should not multiply the probability. Please rewrite the display.
  4. [Fig. 4] The caption uses 'DKL wo.' which is ambiguous. Clarify which configuration each bar denotes (e.g., 'without KL', 'without alignment loss', 'without distribution heads').
  5. [Fig. 5] Selecting S=8 based on test-set performance, without a held-out validation set, is a form of test-set tuning. At minimum, acknowledge this and show validation-set behavior.
  6. [Throughout] Typos and small errors: 'illustracted', 'multi-folded', 'Kullback-Leipner', 'publically', 'all modells', 'form'; Table 1 uses inconsistent decimal places (76.19 vs 96.3); Fig. 6 caption is incomplete ('sheared+rotoscaledcanonicalized by').

Circularity Check

0 steps flagged

No circularity found; empirical evaluation is self-contained.

full rationale

The paper is an empirical method-evaluation study whose central claim (consistent robustness improvement over other STNs) is supported by Table 1 on external benchmarks (EU-Moth, ECU-Moth) against fixed baselines (STN, Head-STN, Polar, LogPolar, Diffeo-STN, P-STN). The only supervision drawn from the augmentation process is the component-wise alignment loss (eq. 12–13), which uses the known augmentation parameters of the spatially augmented training set as regression targets; this is a standard supervised training signal, not a fitted constant that is later reused as a test-set prediction. At inference, the model must infer transformations from the image, and the test sets are independently constructed by augmenting the original test data. The self-citations [16,17] (both by the present authors) appear only as examples of existing pseudo-canonicalization approaches in the introduction; they are not used to justify the method, uniqueness, or any theoretical constraint, so they are not load-bearing. Notably, Sec. 5.1 abandons the KL term, showing that the variational-prior component is not essential to the reported result. A correctness concern exists in Eq. (7): with the stated λ_s=0.25, the scale domain is [1−1/λ_s,1]=[−3,1], contradicting the claim that λ_s∈(0,1) 'ensures strict positiveness of s', and h∈[−4π,4π] is not a reasonable shear bound. However, this is an internal inconsistency/possible typo, not a circular derivation. Since no prediction reduces by construction to a fitted input or to a self-citation chain, the circularity score is 0.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The model introduces no new physical or mathematical entities. The free parameters are hand-tuned bounds, loss weights, and the sample count S. The main load-bearing axioms are the fixed-order affine decomposition and the availability of augmentation labels matching that decomposition.

free parameters (7)
  • λθ (rotation bound) = 1
    Eq. (7); chosen by hand to allow full-orbit rotations.
  • λs (scale bound) = 0.25
    Eq. (7); chosen by hand to bound scaling to a reasonable range.
  • λh (shear bound) = 0.25
    Eq. (7); chosen by hand to bound shearing to a reasonable range.
  • αT, βT (log-variance bounds) = αT=4, βT=6
    Eq. (8); chosen by hand to bound the variance range.
  • Alignment loss weights (λθ, λs, λh in eq. 12) = λθ=2, λs=λh=0.5
    Chosen by hand; rotation weighted higher as it is more challenging.
  • Sample count S = 8
    Selected from ablation in fig. 5; performance peaks at S=8.
  • Augmentation ranges for shearing, scaling, rotation = not specified
    The training augmentation in section 5.2 is described only as 'random shearing, scaling, and rotation', with no ranges given.
axioms (6)
  • domain assumption Affine matrix M can be decomposed as M = R(θ)·S(sx,sy)·H(hx,hy)·F(rx,ry) in a fixed order with bounded parameters
    Eq. (3); not all 2x2 invertible matrices can be represented with the imposed component bounds; the model restricts to this family, which may exclude some real distortions.
  • domain assumption Translation and reflection are irrelevant for the target application; only rotation, scaling, shearing need canonicalization
    Section 2 'Target Subgroups'; the evaluation and alignment loss only cover these three components, so the method cannot handle residual localization errors or mirror images.
  • domain assumption The frozen tokenizer's patch tokens contain sufficient pose information for the shallow 2-layer localization encoder to regress rotation, scale, and shear
    Section 4 'Shared Tokenizer' and 'Localization Encoder'; if tokenization discards absolute pose, the canonicalizer cannot recover it.
  • ad hoc to paper The augmentation parameters used in training are exactly the ground-truth components of the model's fixed-order decomposition
    Eq. (12) alignment loss; the paper does not specify the augmentation order, so the per-component targets may not correspond to the predicted components.
  • ad hoc to paper Sampling from an unregularized Gaussian posterior at inference approximates the predictive distribution in eq. (16)
    Eq. (17) uses MC samples from pφ(T|x) without the KL term; this is a heuristic approximation, not a justified variational bound.
  • standard math Orbit-stabilizer theorem and group action definitions from Section 2
    Used to motivate orbit-space test construction and estimate orbit sizes.

pith-pipeline@v1.3.0-alltime-deepseek · 12932 in / 15966 out tokens · 176977 ms · 2026-08-04T16:53:07.440177+00:00 · methodology

0 comments
read the original abstract

Fine-grained visual classification (FGVC) remains highly sensitive to geometric variability, where objects appear under arbitrary orientations, scales, and perspective distortions. While equivariant architectures address this issue, they typically require substantial computational resources and restrict the hypothesis space. We revisit Spatial Transformer Networks (STNs) as a canonicalization tool for transformer-based vision pipelines, emphasizing their flexibility, backbone-agnostic nature, and lack of architectural constraints. We propose a probabilistic, component-wise extension that improves robustness. Specifically, we decompose affine transformations into rotation, scaling, and shearing, and regress each component under geometric constraints using a shared localization encoder. To capture uncertainty, we model each component with a Gaussian variational posterior and perform sampling-based canonicalization during inference.A novel component-wise alignment loss leverages augmentation parameters to guide spatial alignment. Experiments on challenging moth classification benchmarks demonstrate that our method consistently improves robustness compared to other STNs.

Figures

Figures reproduced from arXiv: 2509.11218 by Johann Schmidt, Sebastian Stober.

Figure 1
Figure 1. Figure 1: We propose a pseudo-canonicalization frame￾work based on spatial transformers that can be embed￾ded into the training and inference pipelines of vision transformers. Our canonicalization module ϕ is trained in an end-to-end fashion with the pre-trained classifier backbone. The pre-trained frozen tokenizer is used to map the transformed input image to visual tokens, which are used to regress an affine trans… view at source ↗
Figure 2
Figure 2. Figure 2: We leverage the frozen tokenizer of the backbone to obtain patch tokens, which are processed by a localization encoder into a shared representation. Separate regression heads predict parameters for rota￾tion, scaling, and shearing, which are composed into an affine matrix Tˆ . The input image is rectified via Tˆ −1 , re-tokenized (with tied weights), and passed to the down￾stream vision transformer for cla… view at source ↗
Figure 3
Figure 3. Figure 3: Samples from the EU-Moths [1] and ECU￾Moths [32] FGVC datasets. In practice we approximate this expectation with S Monte Carlo samples, pϕ(y | x) ≈ 1 S X S s=1 p  y | ρ(Tˆ s) −1x  , (17) where Tˆ s ∼ pϕ(T | ˆ x). This enforces approximate affine invariance by marginalizing transformation uncertainty during both training and inference. 5 Experiments We train and evaluate multiple baselines across sev￾eral… view at source ↗
Figure 4
Figure 4. Figure 4: Top-1 test accuracy decrease on the vanilla and roto-scaled EU-Moth benchmark. These ablation results highlight the performance drops when disabling or replacing parts of our algorithm. padding to prevent boundary artifacts during spa￾tial transformations. Datasets We focus our experiments on moth FGVC benchmarks (for biodiversity monitoring sys￾tems): Ecuador-Moth [32] comprises 1445 samples and 675 class… view at source ↗
Figure 6
Figure 6. Figure 6: ϕ learned to zoom in and align moths to some extend. performance. Most significantly, we found that the decomposed regression approach in eq. (8) (com￾pared to direct regression of Tˆ as in [18]) yielded the largest performance gain. The probabilistic re￾gression of T in eq. (9), rather than deterministic re￾gression, produced the second-largest improvement. We also evaluated the impact of our token-based … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 13 canonical work pages · 7 internal anchors

  1. [1]

    Deep Learning Pipeline for Automated Vi- sual Moth Monitoring: Insect Localization and Species Classification

    D. Korsch, P. Bodesheim, and J. Denzler. “Deep Learning Pipeline for Automated Vi- sual Moth Monitoring: Insect Localization and Species Classification”. In:INFORMATIK (2021).doi: 10.18420/informatik2021-036

    work page doi:10.18420/informatik2021-036 2021

  2. [2]

    Robustness through Data Augmen- tation Loss Consistency

    T. Huang, S. A. Halbe, C. Sankar, P. Amini, S. Kottur, A. Geramifard, M. Razaviyayn, and A. Beirami. “Robustness through Data Augmen- tation Loss Consistency”. In:Transactions on Machine Learning Research (TMLR)(2023). doi:10.48550/arXiv.2110.11205

    Pith/arXiv arXiv doi:10.48550/arxiv.2110.11205 2023

  3. [3]

    Naturally Occurring Equivari- ance in Neural Networks

    C. Olah, N. Cammarata, C. Voss, L. Schubert, and G. Goh. “Naturally Occurring Equivari- ance in Neural Networks”. In:Distill(2020). doi:10.23915/distill.00024.004

    work page doi:10.23915/distill.00024.004 2020

  4. [4]

    A Group-Theoretic Framework for Data Augmentation

    S. Chen, E. Dobriban, and J. Lee. “A Group- Theoretic Framework for Data Augmentation”. In:Advances in Neural Information Processing Systems (NeurIPS). 2020.doi: 10 . 48550 / arXiv.1907.10905

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1907.10905 2020

  5. [5]

    A Kernel Theory of Modern Data Augmentation

    T. Dao, A. Gu, A. Ratner, V. Smith, C. De Sa, and C. Re. “A Kernel Theory of Modern Data Augmentation”. In:International Conference on Machine Learning (ICML). 2019.doi: 10. 48550/arXiv.1803.06084

    Pith/arXiv arXiv doi:10.48550/arxiv.1803.06084 2019

  6. [6]

    The Effects of Regularization and Data Augmen- tation are Class Dependent

    R. Balestriero, L. Bottou, and Y. LeCun. “The Effects of Regularization and Data Augmen- tation are Class Dependent”. In:Conference on Neural Information Processing Systems (NeurIPS)(2022).doi: 10 . 5555 / 3600270 . 3603015

    2022

  7. [7]

    Enhanced Yolov8 net- work with Extended Kalman Filter for wildlife detection and tracking in complex environ- ments

    L. Jiang and L. Wu. “Enhanced Yolov8 net- work with Extended Kalman Filter for wildlife detection and tracking in complex environ- ments”. In:Ecological Informatics(2024).doi: 10.1016/j.ecoinf.2024.102856

    arXiv 2024

  8. [8]

    Making Convolutional Networks Shift-Invariant Again

    R. Zhang. “Making Convolutional Networks Shift-Invariant Again”. In:International Con- ference on Machine Learning (ICML). 2019. doi:10.48550/arXiv.1904.11486

    Pith/arXiv arXiv doi:10.48550/arxiv.1904.11486 2019

  9. [9]

    Group Equivariant Convolutional Networks

    T. Cohen and M. Welling. “Group Equivariant Convolutional Networks”. In:International Conference on Machine Learning (ICML). 2016.doi:10.5555/3045390.3045705

    arXiv 2016

  10. [10]

    Steerable CNNs

    T. S. Cohen and M. Welling. “Steerable CNNs”. In:International Conference on Learning Representations (ICLR). 2017.doi: 10.48550/arXiv.1612.08498

    Pith/arXiv arXiv doi:10.48550/arxiv.1612.08498 2017

  11. [11]

    Meta- Learning Symmetries by Reparameterization

    A. Zhou, T. Knowles, and C. Finn. “Meta- Learning Symmetries by Reparameterization”. In:International Conference on Learning Rep- resentations (ICLR)(2020).doi: 10.48550/ arXiv.2007.02933

    Pith/arXiv arXiv doi:10.48550/arxiv.2007.02933 2020

  12. [12]

    Frame Averaging for Invariant and Equivariant Net- work Design

    O. Puny, M. Atzmon, E. J. Smith, I. Misra, A. Grover, H. Ben-Hamu, and Y. Lipman. “Frame Averaging for Invariant and Equivariant Net- work Design”. In:International Conference on Learning Representations. 2022

    2022

  13. [13]

    Ob- ject Orientation Agnosia: A Failure to Find the Axis?

    I. M. Harris, J. A. Harris, and D. Caine. “Ob- ject Orientation Agnosia: A Failure to Find the Axis?” In:Journal of Cognitive Neuroscience (2001).doi:10.1162/08989290152541467

    work page doi:10.1162/08989290152541467 2001

  14. [14]

    Equivariance with Learned Canonicalization Functions

    S.-O. Kaba, A. K. Mondal, Y. Zhang, Y. Ben- gio, and S. Ravanbakhsh. “Equivariance with Learned Canonicalization Functions”. In:In- ternational Conference on Machine Learning (ICML). 2023.doi: 10.48550/arXiv.2211. 06489

    doi:10.48550/arxiv.2211 2023

  15. [15]

    Equivariant Adaptation of Large Pretrained Models

    A. K. Mondal, S. S. Panigrahi, S. -O. Kaba, S. Rajeswar, and S. Ravanbakhsh. “Equivariant Adaptation of Large Pretrained Models”. In: Conference on Neural Information Processing Systems (NeurIPS). 2023.doi: 10 . 48550 / arXiv.2310.01647

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2310.01647 2023

  16. [16]

    Learning Continuous Rotation Canonicalization with Radial Beam Sampling

    J. Schmidt and S. Stober.Learning Contin- uous Rotation Canonicalization with Radial Beam Sampling. 2023.doi: 10.48550/arXiv. 2206.10690

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2206.10690 2023

  17. [17]

    Tilt your Head: Activating the Hidden Spatial-Invariance of Classifiers

    J. Schmidt and S. Stober. “Tilt your Head: Ac- tivating the Hidden Spatial-Invariance of Clas- sifiers”. In:International Conference on Ma- chine Learning (ICML). 2024.doi: 10.48550/ arXiv.2405.03730

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2405.03730 2024

  18. [18]

    Spatial Transformer Net- works

    M. Jaderberg, K. Simonyan, A. Zisserman, and K. Kavukcuoglu. “Spatial Transformer Net- works”. In:International Conference on Neu- ral Information Processing Systems (NeurIPS). 2015.doi:10.48550/arXiv.1506.02025

    Pith/arXiv arXiv doi:10.48550/arxiv.1506.02025 2015

  19. [19]

    Probabilistic Spatial Transformer Networks

    P. Schw¨ obel, F. R. Warburg, M. Jørgensen, K. H. Madsen, and S. Hauberg. “Probabilistic Spatial Transformer Networks”. In:Confer- ence on Uncertainty in Artificial Intelligence (UAI). 2022.doi: 10 . 48550 / arXiv . 2004 . 03637

    2022

  20. [20]

    Learning Un- supervised Parameter-Specific Affine Trans- formation for Medical Images Registration

    X. Chen, Y. Meng, Y. Zhao, R. Williams, S. R. Vallabhaneni, and Y. Zheng. “Learning Un- supervised Parameter-Specific Affine Trans- formation for Medical Images Registration”. In:Medical Image Computing and Computer Assisted Intervention (MICCAI). 2021.doi: 10.1007/978-3-030-87202-1_3

    work page doi:10.1007/978-3-030-87202-1_3 2021

  21. [21]

    Hall.Lie Groups, Lie Algebras, and Repre- sentations: An Elementary Introduction

    B. Hall.Lie Groups, Lie Algebras, and Repre- sentations: An Elementary Introduction. 2015. doi:10.1007/978-1-4614-7116-5_16

    doi:10.1007/978-1-4614-7116-5_16 2015

  22. [22]

    Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges

    M. M. Bronstein, J. Bruna, T. Cohen, and P. Veliˇ ckovi´ c. “Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges”. In: ArXiv(2021).doi: 10.48550/arXiv.2104. 13478. 9

    doi:10.48550/arxiv.2104 2021

  23. [23]

    Understanding when spatial transformer net- works do not support invariance, and what to do about it

    L. Finnveden, Y. Jansson, and T. Lindeberg. “Understanding when spatial transformer net- works do not support invariance, and what to do about it”. In:International Conference on Pattern Recognition (ICPR)(2020).doi: 10.1109/ICPR48806.2021.9412997

    arXiv 2020

  24. [24]

    Foveated convolutions: improving spa- tial transformer networks by modelling the retina

    E. W. A. Harris, M. Niranjan, and J. Hare. “Foveated convolutions: improving spa- tial transformer networks by modelling the retina”. In:Shared Visual Representations in Human and Machine Intelligence: NeurIPS Workshop. 2019.doi: eprints.soton.ac.uk/ id/eprint/441204

    2019

  25. [25]

    Efficient Rotation Invariance in Deep Neural Networks through Artificial Mental Rotation

    L. Tuggener, T. Stadelmann, and J. Schmid- huber. “Efficient Rotation Invariance in Deep Neural Networks through Artificial Mental Ro- tation”. In: (2023).doi: 10 . 48550 / arXiv . 2311.08525

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2311.08525 2023

  26. [26]

    Polar Transformer Networks

    C. Esteves, C. Allen-Blanchette, X. Zhou, and K. Daniilidis. “Polar Transformer Networks”. In:International Conference on Learning Rep- resentations (ICLR). 2018.doi: 10.48550/ arXiv.1709.01889

    Pith/arXiv arXiv doi:10.48550/arxiv.1709.01889 2018

  27. [27]

    Equivari- ant Transformer Networks

    K. S. Tai, P. Bailis, and G. Valiant. “Equivari- ant Transformer Networks”. In:International Conference on Machine Learning (ICML). 2019.doi:10.48550/arXiv.1901.11399

    Pith/arXiv arXiv doi:10.48550/arxiv.1901.11399 2019

  28. [28]

    Inverse Compositional Spatial Transformer Networks

    C.-H. Lin and S. Lucey. “Inverse Compo- sitional Spatial Transformer Networks”. In: Conference on Computer Vision and Pattern Recognition (CVPR). 2017.doi: 10.48550/ arXiv.1612.03897

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1612.03897 2017

  29. [29]

    Deep Diffeomorphic Transformer Networks

    N. S. Detlefsen, O. Freifeld, and S. Hauberg. “Deep Diffeomorphic Transformer Networks”. In:Conference on Computer Vision and Pat- tern Recognition (CVPR)(2018).doi: 10 . 1109/CVPR.2018.00463

    arXiv 2018

  30. [30]

    A Refined Spatial Transformer Network

    C. Shu, X. Chen, C. Yu, and H. Han. “A Refined Spatial Transformer Network”. In:In- ternational Conference on Neural Information Processing (NeurIPS). 2018.doi: 10.1007/ 978-3-030-04182-3_14

    2018

  31. [31]

    Auto- Encoding Variational Bayes

    D. P. Kingma and M. Welling. “Auto- Encoding Variational Bayes.” In:Interna- tional Conference on Learning Representations (ICLR). 2014.doi: 10.48550/arXiv.1312. 6114

    work page doi:10.48550/arxiv.1312 2014

  32. [32]

    Fine-grained Recognition Datasets for Biodiversity Analysis

    E. Rodner, M. Simon, G. Brehm, S. Pietsch, J. W. W¨ agele, and J. Denzler. “Fine-grained Recognition Datasets for Biodiversity Anal- ysis”. In:CVPR Fine-Grained Visual Clas- sification Workshop(2015).doi: 10.48550/ arXiv.1507.00913

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1507.00913 2015

  33. [33]

    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows

    Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo. “Swin Transformer: Hierarchical Vision Transformer using Shifted Windows”. In:2021 IEEE/CVF International Conference on Computer Vision (ICCV). 2021, pp. 9992–10002.doi: 10 . 1109 / ICCV48922 . 2021.00986

    arXiv 2021

  34. [34]

    Imagenet: A large-scale hierarchical image database

    J. Deng, W. Dong, R. Socher, L. -J. Li, K. Li, and L. Fei-Fei. “Imagenet: A large-scale hierarchical image database”. In:Conference on Computer Vision and Pattern Recognition (CVPR). 2009.doi: 10 . 1109 / CVPR . 2009 . 5206848

    2009

  35. [35]

    Decoupled Weight Decay Regularization

    I. Loshchilov and F. Hutter. “Decoupled Weight Decay Regularization”. In:Interna- tional Conference on Learning Representations (ICLR). 2017.doi: 10.48550/arXiv.1711. 05101

    doi:10.48550/arxiv.1711 2017

  36. [36]

    SGDR: Stochas- tic Gradient Descent with Warm Restarts

    I. Loshchilov and F. Hutter. “SGDR: Stochas- tic Gradient Descent with Warm Restarts”. In:5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceed- ings. OpenReview.net, 2017.url: https:// openreview.net/forum?id=Skq89Scxx

    2017

  37. [37]

    Gaussian Error Linear Units (GELUs)

    D. Hendrycks and K. Gimpel. “Gaussian Error Linear Units (GELUs)”. In:ArXiv(2016)

    2016

  38. [38]

    Batch Normaliza- tion: Accelerating Deep Network Training by Reducing Internal Covariate Shift

    S. Ioffe and C. Szegedy. “Batch Normaliza- tion: Accelerating Deep Network Training by Reducing Internal Covariate Shift”. In:In- ternational Conference on Machine Learn- ing (ICML). 2015.doi: 10 . 5555 / 3045118 . 3045167

    2015

  39. [39]

    D. S. Dummit and R. M. Foote.Abstract Al- gebra. 2004.doi:10.1515/9783110691160

    work page doi:10.1515/9783110691160 2004

  40. [40]

    beta-V AE: Learning Basic Vi- sual Concepts with a Constrained Variational Framework

    I. Higgins, L. Matthey, A. Pal, C. P. Burgess, X. Glorot, M. M. Botvinick, S. Mohamed, and A. Lerchner. “beta-V AE: Learning Basic Vi- sual Concepts with a Constrained Variational Framework.” In:International Conference on Learning Representations (ICLR). 2017.doi: openreview.net/forum?id=Sy2fzU9gl

    2017

  41. [41]

    Cyclical Annealing Schedule: A Simple Approach to Mitigating KL Vanishing

    H. Fu, C. Li, X. Liu, J. Gao, A. Celikyilmaz, and L. Carin. “Cyclical Annealing Schedule: A Simple Approach to Mitigating KL Vanishing”. In:Conference of the North American Chapter of the Association for Computational Linguis- tics (NAACL).doi:10.18653/v1/N19-1021

    doi:10.18653/v1/n19-1021

  42. [42]

    Transformations Based on Continuous Piecewise-Affine Velocity Fields

    O. Freifeld, S. Hauberg, K. Batmanghelich, and J. W. Fisher. “Transformations Based on Continuous Piecewise-Affine Velocity Fields”. In:IEEE Transactions on Pattern Analysis and Machine Intelligence(2017).doi: 10 . 1109/TPAMI.2016.2646685. 10

    arXiv 2017