Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

CLIPSym: Delving into Symmetry Detection with CLIP

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A pre-trained CLIP model, prompted with frequent object-name text and decoded through a rotation-equivariant network, detects reflection axes and rotation centers in natural images, beating the previous state of the art on DENDI, SDRW, and

desk verdict CLIP pretraining carries the SOTA result; prompt tuning is a disclosed-but-unquantified caveat, and the equivariance proof needs a scope note. read the letter →

arxiv 2508.14197 v1 pith:7LWIGCYN submitted 2025-08-19 cs.CV

Tinghan Yang , Md Ashiqur Rahman , Raymond A. Yeh This is my paper · ORCID
classification cs.CV
keywords symmetrydetectionCLIPvision-languagemodelssemantic-awarepromptgroupingrotationequivariancesteerablegroupconvolutionreflection
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Symmetry detection has been stuck on a data problem: dedicated models are trained on small, expensive annotations, and accuracy has plateaued. This paper claims that the semantic knowledge already stored in CLIP, a vision-language model trained on hundreds of millions of image–caption pairs, can be redirected at this geometric task. Its system, CLIPSym, writes 25 text prompts made of frequent object names, uses them to modulate CLIP's image features, and decodes the result with a rotation-equivariant network, reporting the best F1 scores on all three standard benchmarks — 66.5% vs 61.7% for reflection and 25.1% vs 22.0% for rotation on DENDI. If the claim holds, symmetry detection stops being an isolated, data-hungry problem and becomes another skill that general web-scale models can be prompted to perform.

What carries the argument

The machinery has three parts. (1) Semantic-Aware Prompt Grouping (SAPG): a fixed set of M prompts, each a space-joined string of K frequent object classes mined from the target dataset and encoded by CLIP's text encoder, identical for every image so they form a stable semantic anchor. (2) A conditioning block: a FiLM layer scales and shifts CLIP's patch features for each prompt, a Transformer models spatial dependencies, and a learned weighted average merges the M conditioned feature sets. (3) A rotation-equivariant upsampler: the aggregated map is lifted to the C8 roto-translation group and processed by steerable G-convolutions. Claim 1, the load-bearing theoretical result, states the deco

What would settle it

Rebuild the SAPG prompt vocabulary using only the DENDI training split, or an unrelated object list such as COCO class names, and rerun the evaluation: if the 66.5 reflection F1 collapses toward EquiSym's 61.7, the reported gain came from prompts informed by the test set rather than from a transferable semantic prior. A secondary check: rotate a test image by exactly 45 degrees and compare the predicted heatmap with the heatmap of the original image rotated by 45 degrees — a mismatch would falsify the unproven C8 equivariance.

Watch

Extended reading notes

Core claim

CLIPSym is a text-conditioned dense predictor: CLIP's patch features are modulated per prompt by a FiLM layer (25 fixed prompts, each four frequent object names), passed through a Transformer, aggregated, and decoded by steerable G-convolutions on the C8 roto-translation group into a symmetry heatmap. The central claim is that this combination beats specialist equivariant networks — 66.5% vs 61.7% reflection F1 and 25.1% vs 22.0% rotation F1 on DENDI — and that ablations attribute the gain to CLIP's internet-scale pretraining, to the object-name prompts, and to the equivariant decoder. The paper also argues that fixed prompts act as a universal semantic anchor for the abstract concept of sym

Load-bearing premise

The result hinges on the assumption that the 25 object-name prompts mined from the DENDI images are a universal semantic prior and not benchmark-specific tuning — the paper does not state whether the test split was excluded from the mining — and, separately, on an equivariance guarantee proven only for 90-degree rotations while the decoder uses 45-degree steps.

Editorial extensions

If this is right

  • Symmetry detection stops depending on large symmetry-specific datasets: a general CLIP backbone with fixed object-name prompts transfers across DENDI, SDRW, and LDRS and beats the specialist baselines.
  • Semantic conditioning and geometric equivariance compose: the full model beats each ingredient alone, so text cues and equivariant inductive biases add rather than conflict.
  • Improvements in web-scale vision-language backbones carry over directly — MetaCLIP pushes reflection F1 to 66.7 — so the method inherits future gains without architectural changes.
  • Because one architecture covers both reflection and rotation, application-level pipelines (image matching, editing, scene understanding) get a single symmetry detector where the current best rotation-only method requires extra 2D-to-3D label conversion and cannot detect reflection axes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The prompt vocabulary is mined from DENDI itself, and the paper never states that the test split was excluded; I would want a run with prompts built from held-out object lists (e.g., COCO classes) before treating the 4.8-point reflection gain as fully generic.
  • The equivariance proof stops at C4 while the decoder uses 45-degree filters with bilinear interpolation; empirically I expect small but nonzero equivariance error at exactly 45 degrees, measurable by comparing heatmaps on 45-degree-rotated inputs.
  • The paper's own theoretical argument for language (Claim 2) is constructive — FiLM could add back missing cues — rather than descriptive; the real evidence for language is the ablation gap, so the mechanism's inner contribution remains partly open.
  • The recipe is not symmetry-specific: semantic prompts plus FiLM conditioning plus an equivariant decoder could plausibly transfer to other dense geometric tasks, such as vanishing-point or horizon estimation, where object semantics and geometry interact.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes CLIPSym, a symmetry-detection model that fine-tunes pre-trained CLIP image and text encoders and feeds the resulting tokens into a FiLM-conditioned Transformer plus a rotation-equivariant (G-convolution) decoder to predict per-pixel reflection-axis and rotation-center heatmaps. A new prompting technique, SAPG, aggregates a fixed set of multi-word prompts built from frequent object classes extracted from the DENDI dataset. The paper reports state-of-the-art F1 scores on DENDI, SDRW, and LDRS (e.g., 66.5% reflection F1 and 25.1% rotation F1 on DENDI), and provides ablations showing contributions from CLIP pretraining, text conditioning, and the equivariant decoder, as well as improved robustness/consistency under rotations. The code is publicly released.

Significance. If the central claims hold, this is a useful first demonstration that CLIP representations can be adapted to symmetric dense prediction, and the reported gains are consistent across three standard datasets. The paper is thorough in ablating the main components (pretraining, text prompts, equivariant design, CLIP variants), and releasing code is a clear strength. However, the empirical SOTA claim currently rests on two issues that must be resolved: the SAPG prompt set is constructed from DENDI without a stated guarantee that the test split was excluded, and the baseline comparison mixes original published numbers with lower reproduced numbers. There is also a direct tension between the abstract/conclusion's unrestricted SOTA claim and the appendix's admission that a recent rotation-only method achieves higher rotation F1 on DENDI. These issues are fixable but are load-bearing for the paper's headline contribution.

major comments (4)
  1. [Appendix A1 / Table 1 / Table 5] The SAPG prompt set is built from frequent object classes 'from the DENDI dataset' using Grounded-SAM, but the manuscript never states whether the train/val/test split was respected. Because the same prompt set is used for all images and Table 5 shows that prompt content changes reflection F1 from 64.4 to 66.5, use of test images in prompt selection would embed test-set statistics into the model and materially inflate the reported DENDI gains. Please state explicitly that prompts were generated from the training split only; if that is already the case, add a sentence saying so. If it is not the case, re-run with train-only prompts and report both results. The same clarification is needed for the hyperparameter selection reported in Tables 5, A3, and A3.6 (M=25, K=4, C8): were these chosen on the validation split or on the test set?
  2. [Sec. 4.2, Table 1] The paper lists both the original EquiSym numbers (64.5/22.5) and its own reproduction (61.7±0.6/22.0±0.7), but the results text says CLIPSym outperforms 'EquiSym∗' by 2.0/2.6, i.e., it uses the original numbers. Against the reproduction the gains are 4.8/3.1. The discrepancy between 64.5 and 61.7 is never explained. If the reproduction follows a different evaluation protocol, then comparing to EquiSym∗ is inconsistent; if it is due to implementation details, the original numbers should not be used as the baseline. Please provide one consistent baseline table and explain the gap.
  3. [Abstract / Conclusion vs Appendix A3.8] The abstract and conclusion state that CLIPSym 'outperforms the current state-of-the-art' and 'achieves state-of-the-art performance for both reflection and rotation symmetry detection.' Appendix A3.8, however, concedes that Seo and Cho [49] achieve a higher F1 score for rotation center detection on DENDI. Unless the SOTA claim is explicitly limited to unified methods that handle both reflection and rotation in one framework, the claim is inconsistent with the appendix. Please qualify the claim and either include the comparison with [49] in the main table or explain why it is not considered a SOTA baseline for the reported setting.
  4. [Sec. 3.1, Claim 1] The theoretical equivariance guarantee is explicitly for C4 (Eq. 12), while the deployed upsampler uses Cn with n=8, as stated in Sec. 3.1 and evaluated in Table A3. The proof of step ③ only shows that C4 follows from Cn equivariance. Since the method is described as having 'theoretical guarantees for rotation equivariance' and the C8 choice is empirically best, please either prove C8 equivariance (including the behavior of bilinear interpolation at 45°) or explicitly state that the guarantee is C4 and that C8 is an empirical design choice.
minor comments (5)
  1. [Sec. 2 (Related work)] The sentence 'Local feature descriptors, such as SIFT [51]' attributes SIFT to reference [51], which is the EquiSym paper. The SIFT citation appears misplaced.
  2. [Appendix A2.2, Eq. (A19)-(A29)] Claim 2's proof constructs β(zt)=δ*(t), where δ*(t) is exactly the additive offset assumed in Hypothesis 1. The conclusion is therefore a direct consequence of the hypothesis and should be labeled as a conditional illustration, not as an independent proof that language helps.
  3. [Table 4 / text] The text says PMCNet has '167.7 FLOPs' while the table caption uses GFLOPs. This should be corrected to GFLOPs for consistency.
  4. [Figure A2 caption] Typo: 'propmpting' should be 'prompting'.
  5. [Appendix A2.2] Typo: 'correpsondingly' should be 'correspondingly'.

Circularity Check

1 steps flagged · score 3.0 of 10

Empirical SOTA claim is self-contained; the only clear circular step is the Appendix 'language helps' proof, which assumes the offset it then adds back.

  1. self definitional [Appendix A2.2, 'A theoretical perspective on the benefits of language', Eq. (A19)-(A30), Claim 2]
    "Provided that language contains cues about symmetry, we assume the best visual features are offset by an additive term δ∗(t) that depends on a language prompt t, plus zero-mean noise εI: ZI = Z∗I − δ∗(t) + εI ... Then we make a claim that language is beneficial under Hypothesis 1: ... Simply choose γ(zt) = 1(all ones vector), β(zt) = δ∗(t) (A23) and apply to A20, then ZI|t = ZI + δ∗(t) = Z∗I + εI."

    The claimed benefit of language is built into Hypothesis 1: δ∗(t) is defined as the exact additive gap between the actual CLIP features and the 'best' visual features, and δ∗(t) ≠ 0 is assumed whenever language supplies symmetry cues. The proof then chooses FiLM parameters γ=1 and β=δ∗(t) to add that exact offset back, so E||ZI|t − Z∗I|| < E||ZI − Z∗I|| follows by construction. No independent estimate of δ∗(t), and no argument that CLIP text embeddings can recover it, is provided. The conclusion 'language is beneficial' is therefore the assumption, not a derived consequence.

full rationale

The central claim—state-of-the-art F1 on DENDI, SDRW, and LDRS—is an empirical benchmark result, not a derivation from a fitted equation. The decoder, focal loss, and training protocol are standard, and CLIP pretraining is an external, independently available model. The C4 equivariance proof (Claim 1) is only a proof sketch for C4 while the decoder uses C8 with bilinear interpolation; this is an incompleteness/correctness concern, not circularity. The SAPG prompt set is extracted from DENDI (Appendix A1) and the prompt-structure hyperparameters M,K are searched on DENDI (Appendix A3.6); the paper never states that the test split was excluded from the Grounded-SAM object-class extraction, so the reported SOTA margin could partly reflect benchmark-specific prompt design. That is an evaluation-hygiene concern, but the text does not exhibit a specific reduction from prompt selection to the reported numbers, so it is not scored as circular under the evidence rule. The one genuine tautological step is Appendix A2.2's Claim 2: the assumed offset δ∗(t) already encodes the conclusion that language provides the missing symmetry cue, and the proof chooses β=δ∗(t) to 'add it back'. This step is self-definitional but not load-bearing for the empirical headline. Accordingly, the score is a modest 3.

Assumptions & free parameters 8 free parameters · 7 assumptions · 0 invented entities

The central empirical claim rests on a set of hand-tuned and data-selected hyperparameters, with the prompt set being the most dataset-specific. The theoretical equivariance guarantee relies on standard permutation/group equivariance results plus an unproven assertion about interpolation equivariance for C8. No new physical or conceptual entities are introduced.

free parameters (8)
  • focal loss balance factor α = 0.85 (reflection), 0.95 (rotation)
    Set by hand based on class imbalance; Section A4.
  • focal loss focusing parameter λ = 2.0
    Standard setting, Section A4.
  • number of prompts M = 25
    Selected via hyperparameter search over {1,10,25,50}; Tables A4 and A3.6.
  • words per prompt K = 4
    Selected from {1,4,8,16,32} on DENDI validation; Table A4.
  • equivariance group size = C8
    Selected by ablation over C4, C6, C8, C12, C16 in Table A3.
  • SAPG prompt content = 25 prompts listed in Table A4
    Derived from frequent object classes of the DENDI dataset with Grounded-SAM; data-dependent.
  • learning rate = 1e-5
    Selected from grid search; Appendix A4.
  • decoder channels and transformer layers = channels [64,32,16,1], LB=3
    Architecture choices specified in Appendix A4.
assumptions (7)
  • domain assumption CLIP pre-trained on large-scale image-text data encodes symmetry-related visual cues.
    The central hypothesis, stated in the Introduction and used to justify fine-tuning CLIP for symmetry detection.
  • domain assumption Grounded-SAM reliably extracts frequent object classes from DENDI images.
    Used in Appendix A1 to construct the SAPG prompt set.
  • standard math Transformer blocks are permutation equivariant with respect to token order.
    Used in Claim 1 proof, Step 2, citing Yun et al. [74].
  • standard math FiLM affine transformation is permutation equivariant.
    Used in Claim 1 proof, Step 1.
  • standard math For C4 rotations on a square grid, the rotation action is a permutation of patch locations.
    Used in Claim 1 proof, Step 2.
  • domain assumption The upsampler built from G-Conv layers and bilinear interpolation is equivariant to Cn for n a multiple of 4.
    Asserted in Claim 1, Step 3, without a rigorous proof for n>4; bilinear interpolation is only approximately equivariant to arbitrary rotations.
  • ad hoc to paper Hypothesis 1 in Appendix A2.2: language provides symmetry cues that offset visual features by an additive term.
    This is a hypothetical condition used in the theoretical discussion of why language helps; not needed for the empirical results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLIPSym: Delving into Symmetry Detection with CLIP." pith.science (2026). https://pith.science/paper/7LWIGCYN

@misc{pith2026250814197,
  author       = {Pith},
  title        = {Pith review of: CLIPSym: Delving into Symmetry Detection with CLIP},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7LWIGCYN}},
  note         = {Machine review of arXiv:2508.14197}
}
abstract

Symmetry is one of the most fundamental geometric cues in computer vision, and detecting it has been an ongoing challenge. With the recent advances in vision-language models,~i.e., CLIP, we investigate whether a pre-trained CLIP model can aid symmetry detection by leveraging the additional symmetry cues found in the natural image descriptions. We propose CLIPSym, which leverages CLIP's image and language encoders and a rotation-equivariant decoder based on a hybrid of Transformer and $G$-Convolution to detect rotation and reflection symmetries. To fully utilize CLIP's language encoder, we have developed a novel prompting technique called Semantic-Aware Prompt Grouping (SAPG), which aggregates a diverse set of frequent object-based prompts to better integrate the semantic cues for symmetry detection. Empirically, we show that CLIPSym outperforms the current state-of-the-art on three standard symmetry detection datasets (DENDI, SDRW, and LDRS). Finally, we conduct detailed ablations verifying the benefits of CLIP's pre-training, the proposed equivariant decoder, and the SAPG technique. The code is available at https://github.com/timyoung2333/CLIPSym.

Figures

Figures reproduced from arXiv: 2508.14197 by the authors.

Figure 1
Figure 1. Overview of the proposed CLIPSym architecture. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Visualization of the reflection and rotation symmetry detection on the DENDI dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Examples of the original image, ground truth, rotated ground truth, EquiSym and CLIPSym’s predicted heatmaps [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Classical versus Deep Mirror-Symmetry Scoring: A Benchmark of Thirteen Methods

    cs.CV 2026-07 accept novelty 6.5 of 10

    Tuned HOG nearly matches the best frozen deep features for mirror-symmetry scoring (~0.03 skill gap) at ~300× lower CPU cost; discrimination lives in mid-scale oriented features.

Reference graph

Works this paper leans on

78 extracted references · 2 canonical work pages · cited by 1 Pith paper

  1. [49]

    Seo and M

    A. Seo and M. Cho. Leveraging 3d geometric priors in 2d rotation symmetry detection. In CVPR, 2025. 1, 17

    2025

  2. [1]

    I. R. Atadjanov and S. Lee. Reflection symmetry detec- tion via appearance of structure descriptor. In ECCV,

  3. [2]

    Bahng, A

    H. Bahng, A. Jahanian, S. Sankaranarayanan, and P. Isola. Exploring visual prompts for adapting large- scale models. arXiv preprint arXiv:2203.17274, 2022. 2

    arXiv 2022

  4. [3]

    Cicconet, V

    M. Cicconet, V . Birodkar, M. Lund, M. Werman, and D. Geiger. A convolutional approach to reflection sym- metry. Pattern Recognition Letters, 2017. 1

    2017

  5. [4]

    Cohen and M

    T. Cohen and M. Welling. Group equivariant convolu- tional networks. In ICML, 2016. 2, 4

    2016

  6. [5]

    T. S. Cohen and M. Welling. Steerable CNNs. InICLR,

  7. [6]

    J. D. Delius and G. Habers. Symmetry: can pigeons conceptualize it? Behavioral biology, 1978. 1

    1978

  8. [7]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weis- senborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Min- derer, G. Heigold, S. Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 17

    arXiv 2010

Show all 78 references
  1. [8]

    Driver, G

    J. Driver, G. C. Baylis, and R. D. Rafal. Preserved figure-ground segregation and symmetry perception in visual neglect. Nature, 1992. 1

    1992

  2. [9]

    Dumoulin, E

    V . Dumoulin, E. Perez, N. Schucher, F. Strub, H. d. Vries, A. Courville, and Y . Bengio. Feature-wise trans- formations. Distill, 2018. 3

    2018

  3. [10]

    Ecins, C

    A. Ecins, C. Fermüller, and Y . Aloimonos. Cluttered scene segmentation using the symmetry constraint. In ICRA, 2016. 1

    2016

  4. [11]

    Fukushima and M

    K. Fukushima and M. Kikuchi. Symmetry axis ex- traction by a neural network. Neurocomputing, 2006. 2

    2006

  5. [12]

    Funk and Y

    C. Funk and Y . Liu. Beyond planar symmetry: Mod- eling human perception of reflection and rotation sym- metries in the wild. In ICCV, 2017. 1, 2, 5, 6

    2017

  6. [13]

    C. Funk, S. Lee, M. R. Oswald, S. Tsogkas, W. Shen, A. Cohen, S. Dickinson, and Y . Liu. 2017 iccv chal- lenge: Detecting symmetry in the wild. In ICCVW,

    2017

  7. [14]

    Gnutti, F

    A. Gnutti, F. Guerrini, and R. Leonardi. Combining ap- pearance and gradient information for image symmetry detection. IEEE TIP, 2021. 2

    2021

  8. [15]

    Hartford, D

    J. Hartford, D. Graham, K. Leyton-Brown, and S. Ra- vanbakhsh. Deep models of interactions across sets. In Proc. ICML, 2018. 2

    2018

  9. [16]

    D. C. Hauagge and N. Snavely. Image matching using local symmetry features. In CVPR, 2012. 1

    2012

  10. [17]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In ICCV, 2016. 5

    2016

  11. [18]

    X. Hu, C. Zhang, Y . Zhang, B. Hai, K. Yu, and Z. He. Learning to adapt CLIP for few-shot monocular depth estimation. In WACV, 2024. 2

    2024

  12. [19]

    J. Je, J. Liu, G. Yang, B. Deng, S. Cai, G. Wetzstein, O. Litany, and L. Guibas. Robust symmetry detection via riemannian langevin dynamics. In SIGGRAPH Asia, 2024. 1

    2024

  13. [20]

    C. Jia, Y . Yang, Y . Xia, Y .-T. Chen, Z. Parekh, H. Pham, Q. Le, Y .-H. Sung, Z. Li, and T. Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 1

    2021

  14. [21]

    Kawano, W

    M. Kawano, W. Kumagai, A. Sannai, Y . Iwasawa, and Y . Matsuo. Group equivariant conditional neural pro- cesses. arXiv preprint arXiv:2102.08759, 2021. 2

    2021 arXiv

  15. [22]

    Keller and Y

    Y . Keller and Y . Shkolnisky. A signal processing ap- proach to symmetry detection. IEEE TIP, 2006. 2

    2006

  16. [23]

    M. U. Khattak, H. Rasheed, M. Maaz, S. Khan, and F. S. Khan. Maple: Multi-modal prompt learning. In CVPR, 2023. 2

    2023

  17. [24]

    Lee and Y

    S. Lee and Y . Liu. Skewed rotation symmetry group detection. IEEE TPAMI, 2009. 2

    2009

  18. [25]

    Lengyel, O

    A. Lengyel, O. Strafforello, R.-J. Bruintjes, A. Gielisse, and J. van Gemert. Color equivariant convolutional networks. In NeurIPS, 2024. 2

    2024

  19. [26]

    Lin, L.-L

    H.-C. Lin, L.-L. Wang, and S.-N. Yang. Extracting periodicity of a regular texture based on autocorrelation functions. Pattern recognition letters, 1997. 2

    1997

  20. [27]

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Dollár. Focal loss for dense object detection. In ICCV, 2017. 5

    2017

  21. [28]

    I.-J. Liu, R. A. Yeh, and A. G. Schwing. Pic: permuta- tion invariant critic for multi-agent deep reinforcement learning. In Proc. CORL, 2020. 2

    2020

  22. [29]

    I.-J. Liu, Z. Ren, R. A. Yeh, and A. G. Schwing. Se- mantic tracklets: An object-centric representation for visual multi-agent reinforcement learning. In Proc. IROS, 2021. 2

    2021

  23. [30]

    J. Liu, G. Slota, G. Zheng, Z. Wu, M. Park, S. Lee, I. Rauschert, and Y . Liu. Symmetry detection from realworld images competition 2013: Summary and results. In CVPRW, 2013. 1

    2013

  24. [31]

    Y . Liu, H. Hel-Or, C. S. Kaplan, L. Van Gool, et al. Computational symmetry in computer vision and com- puter graphics. Foundations and Trends® in Computer Graphics and Vision, 2010. 1

    2010

  25. [32]

    Lüddecke and A

    T. Lüddecke and A. Ecker. Image segmentation using text and image prompts. In CVPR, 2022. 2

    2022

  26. [33]

    Luká ˇc, D

    M. Luká ˇc, D. S `ykora, K. Sunkavalli, E. Shechtman, O. Jamriška, N. Carr, and T. Pajdla. Nautilus: Re- covering regional symmetry transformations for image editing. ACM TOG, 2017. 1

    2017

  27. [34]

    S. Park, A. Senocak, and J. S. Chung. Can CLIP help sound source localization? In WACV, 2024. 2

    2024

  28. [35]

    H. Pashler. Coordinate frame for symmetry detection and object recognition. Journal of Experimental Psy- chology: Human Perception and Performance, 1990. 1

    1990

  29. [36]

    V . S. N. Prasad and L. S. Davis. Detecting rotational symmetries. In ICCV, 2005. 1, 2

    2005

  30. [37]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 6

    2021

  31. [38]

    M. A. Rahman and R. A. Yeh. Truly scale-equivariant deep nets with fourier layers. In NeurIPS, 2024. 2

    2024

  32. [39]

    M. A. Rahman and R. A. Yeh. Group downsam- pling with equivariant anti-aliasing. In The Thir- teenth International Conference on Learning Repre- sentations, 2025. URL https://openreview. net/forum?id=sOte83GogU. 2

    2025

  33. [40]

    Ramesh, M

    A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. V oss, A. Radford, M. Chen, and I. Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1

    2021

  34. [41]

    Rasheed, M

    H. Rasheed, M. U. Khattak, M. Maaz, S. Khan, and F. S. Khan. Fine-tuned CLIP models are efficient video learners. In CVPR, 2023. 2

    2023

  35. [42]

    Rauschert, J

    I. Rauschert, J. Liu, K. Brockelhurst, S. Kashyap, and Y . Liu. Symmetry detection competition: A summary of how the competition is carried out. InCVPRW, 2011. 1

    2011

  36. [43]

    Ravanbakhsh, J

    S. Ravanbakhsh, J. Schneider, and B. Poczos. Deep learning with sets and point clouds. In Proc. ICLR workshop, 2017. 2

    2017

  37. [44]

    T. Ren, S. Liu, A. Zeng, J. Lin, K. Li, H. Cao, J. Chen, X. Huang, Y . Chen, F. Yan, et al. Grounded sam: As- sembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 12

    2024 arXiv

  38. [45]

    Rodríguez, A

    I. Rodríguez, A. Gumbert, N. Hempel de Ibarra, J. Kunze, and M. Giurfa. Symmetry is in the eye of the ‘beeholder’: innate preference for bilateral symmetry in flower-naïve bumblebees. Naturwissenschaften, 2004. 1

    2004

  39. [46]

    R. A. Rojas-Gomez, T.-Y . Lim, A. Schwing, M. Do, and R. A. Yeh. Learnable polyphase sampling for shift invariant and equivariant convolutional networks. Advances in Neural Information Processing Systems, 35:35755–35768, 2022. 2

    2022

  40. [47]

    R. A. Rojas-Gomez, T.-Y . Lim, M. N. Do, and R. A. Yeh. Making vision transformers truly shift-equivariant. In CVPR, 2024. 2

    2024

  41. [48]

    Schuhmann, R

    C. Schuhmann, R. Vencu, R. Beaumont, R. Kaczmar- czyk, C. Mullis, A. Katta, T. Coombes, J. Jitsev, and A. Komatsuzaki. Laion-400M: Open dataset of CLIP- filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 12

    2021 arXiv

  42. [50]

    A. Seo, W. Shim, and M. Cho. Learning to discover reflection symmetry via polar matching convolution. In ICCV, 2021. 1, 2, 5, 6, 7, 8, 14, 15, 17

    2021

  43. [51]

    A. Seo, B. Kim, S. Kwak, and M. Cho. Reflection and rotation symmetry detection via equivariant learning. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 8, 13, 14, 15, 17

    2022

  44. [52]

    D. Shen, H. H. Ip, and E. K. Teoh. Robust detection of skewed symmetries by combining local and semi-local affine invariants. Pattern Recognition, 2001. 1, 2

    2001

  45. [53]

    S. N. Sinha, K. Ramnath, and R. Szeliski. Detecting and reconstructing 3d mirror symmetric objects. In ECCV, 2012. 1

    2012

  46. [54]

    Sosnovik, M

    I. Sosnovik, M. Szmaja, and A. Smeulders. Scale- equivariant steerable networks. arXiv preprint arXiv:1910.11093, 2019. 2

    1910 arXiv

  47. [55]

    Tsogkas and I

    S. Tsogkas and I. Kokkinos. Learning-based symmetry detection in natural images. In ECCV, 2012. 2

    2012

  48. [56]

    C. W. Tyler. Human symmetry perception and its com- putational analysis. Psychology Press, 2003. 1

    2003

  49. [57]

    Vetter, T

    T. Vetter, T. Poggio, and H. Bülthoff. The importance of symmetry and virtual views in three-dimensional object recognition. Current Biology, 1994. 1

    1994

  50. [58]

    von Fersen, C

    L. von Fersen, C. S. Manos, B. Goldowsky, and H. Roit- blat. Dolphin detection and conceptualization of sym- metry. Marine mammal sensory systems, 1992. 1

    1992

  51. [59]

    Z. Wang, L. Fu, and Y . Li. Unified detection of skewed rotation, reflection and translation symmetries from affine invariant contour features. Pattern recognition,

  52. [60]

    Z. Wang, Z. Tang, and X. Zhang. Reflection symmetry detection using locally affine invariant edge correspon- dence. IEEE TIP, 2015. 2

    2015

  53. [61]

    Weiler and G

    M. Weiler and G. Cesa. General E(2)-equivariant steer- able cnns. In NeurIPS, 2019. 2

    2019

  54. [62]

    Weiler, M

    M. Weiler, M. Geiger, M. Welling, W. Boomsma, and T. S. Cohen. 3D steerable CNNs: Learning rotationally equivariant features in volumetric data. In NeurIPS, 2018

    2018

  55. [63]

    Weiler, F

    M. Weiler, F. A. Hamprecht, and M. Storath. Learning steerable filters for rotation equivariant cnns. In CVPR,

  56. [64]

    Wilder, M

    J. Wilder, M. Rezanejad, S. Dickinson, K. Siddiqi, A. Jepson, and D. B. Walther. Local contour symmetry facilitates scene categorization. Cognition, 2019. 1

    2019

  57. [65]

    Worrall and M

    D. Worrall and M. Welling. Deep scale-spaces: Equiv- ariance over scale. In NeurIPS, 2019. 2

    2019

  58. [66]

    Z. Wu, Y . Liu, H. Dong, X. Tang, J. Yang, B. Jin, M. Chen, and X. Wei. R2det: Exploring relaxed rota- tion equivariance in 2d object detection. In ICLR, 2025. 1

    2025

  59. [67]

    H. Xu, S. Xie, X. E. Tan, P.-Y . Huang, R. Howes, V . Sharma, S.-W. Li, G. Ghosh, L. Zettlemoyer, and C. Feichtenhofer. Demystifying CLIP data. arXiv preprint arXiv:2309.16671, 2023. 7

    2023 arXiv

  60. [68]

    R. Xu, K. Yang, K. Liu, and F. He. E(2)-equivariant vision transformer. In Proc. UAI, 2023. 2

    2023

  61. [69]

    C. Xue, W. Zhang, Y . Hao, S. Lu, P. H. Torr, and S. Bai. Language matters: A weakly supervised vision- language pre-training approach for scene text detection and spotting. In ECCV, 2022. 2

    2022

  62. [70]

    Yeh, Y .-T

    R. Yeh, Y .-T. Hu, and A. Schwing. Chirality nets for hu- man pose regression. Advances in Neural Information Processing Systems, 32, 2019. 2

    2019

  63. [71]

    R. A. Yeh, A. G. Schwing, J. Huang, and K. Murphy. Diverse generation for multi-agent sports games. In Proc. CVPR, 2019. 2

    2019

  64. [72]

    R. A. Yeh, Y .-T. Hu, M. Hasegawa-Johnson, and A. Schwing. Equivariance discovery by learned parameter-sharing. In International Conference on Artificial Intelligence and Statistics, pages 1527–1545. PMLR, 2022. 2

    2022

  65. [73]

    L. Yuan, D. Chen, Y .-L. Chen, N. Codella, X. Dai, J. Gao, H. Hu, X. Huang, B. Li, C. Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1

    2021 arXiv

  66. [74]

    C. Yun, S. Bhojanapalli, A. S. Rawat, S. J. Reddi, and S. Kumar. Are transformers universal approximators of sequence-to-sequence functions? In ICLR, 2020. 4

    2020

  67. [75]

    Zaheer, S

    M. Zaheer, S. Kottur, S. Ravanbakhsh, B. Poczos, R. R. Salakhutdinov, and A. J. Smola. Deep sets. InNeurIPS,

  68. [76]

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer. Sigmoid loss for language image pre-training. InICCV,

  69. [77]

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu. Conditional prompt learning for vision-language models. In CVPR,

  70. [78]

    man pole stand

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu. Learning to prompt for vision-language models. IJCV, 2022. 2 Appendix • In Sec. A1, we introduce more details about the struc- ture of SAPG and provide several examples of the object classes and the set of prompts that are used in our e...

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.