{"id":"6c1e754d-ad41-449f-9447-863781ead736","arxiv_id":"2608.08963","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Fourier self-supervision, using low- and high-pass filtered image views with frequency-specialized latent dimensions, improves fine-grained Generalized Category Discovery over SimGCD and SelEx baselines.","lead":"A method for discovering fine-grained object categories that filters images into low- and high-frequency versions and trains a contrastive model on all three views. It reports large gains on fine-grained datasets like CUB-200 and Aircraft, though gains shrink or reverse on DINOv2 and Stanford Cars.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DINOv2 Stanford-Cars regression is hidden by the paper's own aggregation","rationale":"The reader's weakest_assumption focuses on the untested frequency-category hypothesis that filtered reconstructions preserve category information. This is a real gap, but I see a more concrete and more load-bearing issue: the DINOv2 Stanford-Cars result is a clear counterexample to the abstract's unqualified claim. The claimed 'outperforms state-of-the-art' is based on averages that hide one negative dataset row (-2.1 All, -3.5 Novel) and one non-improvement (Oxford-Pet). Since the paper's own Table 1 shows this, the claim is not internally inconsistent, but it is overstated. The reader's conditional verdict is already appropriate; my concern strengthens the case for requiring the authors to either explain the failure mode or restrict the claim. The frequency-category assumption is also untested, but it is rendered secondary by the observed regression, which is directly visible in the paper's own tables. A focused check on Stanford-Cars DINOv2 with seeds would settle whether the regression is noise or a genuine limitation; either way the paper should report it transparently rather than burying it in an average.","tokens_in":22913,"tokens_out":1790,"duration_ms":14495,"concrete_test":"Recompute Table 1's DINOv2 averages excluding Stanford-Cars: if the headline claim 'outperforms state-of-the-art' is meant to hold across all fine-grained datasets, the Stanford-Cars DINOv2 row (-2.1 All, -3.5 Novel) must either be explained or the claim must be restricted to the settings where it holds. A direct check: run the exact FourEx protocol on Stanford-Cars with DINOv2, repeat with 3 random seeds, and report mean and standard deviation. If the -2.1 All gap reproduces, the central claim should be revised to 'improves on most fine-grained datasets' rather than 'outperforms state-of-the-art.' As a secondary check, measure category-information preservation of the filtered views (e.g., train a linear probe on low-pass and high-pass reconstructions at the chosen SNR and report top-1 accuracy) to test whether the SNR threshold is a valid proxy for category discriminability.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim, in the abstract and Section 5.2, is that Fourier Self-Supervision 'outperforms state-of-the-art methods, even when the number of classes is unknown.' The strongest support is Table 1: FourEx improves over SelEx on DINOv1 by +6.9/+3.3/+2.6 All on CUB/Aircraft/Cars and on DINOv2 by +0.4/+1.7/-2.1 All. The Stanford-Cars DINOv2 row is a negative result (-2.1 All, -3.5 Novel), so the claim rests on the average across datasets. Table 2's Herbarium19 DINOv1 gain (+4.9 All) is listed in the appendix, not in the main tables, and the main text says 'We also report results on Oxford-IIIT Pet' without giving the selected comparison; FourEx actually ties or slightly underperforms SelEx on Oxford-Pet (+0.3 All, -0.3 Known). Thus the aggregate 'outperforms' is carried by CUB and Aircraft (and Herbarium in the appendix), while the Stanford-Cars DINOv2 regression and the Oxford-Pet non-improvement are acknowledged in the main text only as 'the exception of Stanford Cars.' The claimed consistency across backbones is therefore not established. The method's internal motivation (low frequencies for abstraction, high frequencies for fine-grained discrimination, Section 4 and Fig. 2) is also not directly validated: the SNR-15dB threshold is a quality metric, not a category-information metric, and the paper does not measure whether the filtered views preserve category content. However, the stronger and more falsifiable concern is the DINOv2 Stanford-Cars negative result, which undercuts the unqualified 'outperforms state-of-the-art' claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Fourier Self-Supervision (FourEx, with variants FourSim and FourGCD) for fine-grained Generalized Category Discovery. The method applies low-pass and high-pass filters in the Fourier domain to construct alternative views of each image, uses these views in supervised and unsupervised contrastive losses over dedicated portions of the latent dimension, and adds a classification loss on the filtered reconstructions. Experiments on CUB-200, FGVC-Aircraft, Stanford-Cars, Oxford-IIIT Pet, and Herbarium19, with DINOv1 and DINOv2 backbones, report gains over the SelEx and SimGCD baselines, and the paper claims state-of-the-art results including when the number of categories is unknown.","tokens_in":23234,"tokens_out":3293,"duration_ms":33860,"significance":"If the reported gains hold, the method is significant as a simple, plug-and-play self-supervision signal that improves fine-grained category discovery, a setting where existing contrastive GCD methods are known to struggle. The paper's strengths include a clear motivation grounded in the frequency spectrum, ablations of each loss component, hyperparameter sensitivity studies, time-complexity reporting, and a public code release. However, the central 'outperforms state-of-the-art' claim is weakened by the DINOv2 results, where the average All accuracy is exactly tied with the SelEx baseline and Stanford-Cars drops by 2.1 points, and by the absence of error bars or multiple-seed experiments. The method's core assumption that filtered views retain the category information needed for contrastive learning is not directly validated.","major_comments":[{"comment":"The abstract and Section 5.2 claim that Fourier Self-Supervision outperforms state-of-the-art methods, but the DINOv2 rows in Table 1 show an average All accuracy of 83.1 for both SelEx and FourEx, i.e., a +0.0 aggregate gain. Stanford-Cars specifically drops by 2.1 All and 3.5 Novel. Since the claim is based on the average across datasets, a single negative result on one of the three fine-grained benchmarks is load-bearing. The authors should either revise the claim to per-dataset improvements or provide statistical evidence that the aggregate is robust.","section":"Section 5.2, Table 1"},{"comment":"No error bars or multiple-seed results are reported anywhere. GCD evaluations use k-means followed by Hungarian matching, which is sensitive to initialization and seed, and the reported differences with DINOv2 are small (e.g., +0.4 All on CUB, -2.1 on Cars). Without standard deviations over at least three seeds, the reader cannot tell whether these differences are meaningful. Please add mean and standard deviation for the main comparisons, especially for the negative Stanford-Cars result.","section":"Tables 1-5, 8-10"},{"comment":"The method rests on the assumption that an object's category remains recognizable in low- and high-pass filtered versions and that low frequencies carry broad category information while high frequencies carry fine-grained distinctions. The 15 dB SNR threshold is a signal-quality metric, not a measure of category information content. The paper provides no quantitative check that the filtered views preserve enough discriminative signal for the contrastive losses to be beneficial rather than noisy. I suggest a direct measurement, e.g., linear-probe or k-NN accuracy on the filtered reconstructions versus the originals, or an SNR sweep evaluated per dataset on both All and Novel accuracy.","section":"Section 4, Eqs. (6)-(9)"},{"comment":"The loss weights alpha_low, alpha_high, alpha_cls, the SNR threshold, and the Gaussian blur kernel are all selected on CUB-200 with DINOv1, and the same values are then applied across all datasets and backbones. The hyperparameter study in Table 5 is only on CUB. This makes the cross-dataset generalization claims less secure, especially because the SNR cutoff itself is computed per dataset but the loss coefficients are not adapted. Please show sensitivity of the main results to these choices on at least one additional dataset or provide a justification for why the CUB-tuned values transfer.","section":"Section 5.3, Table 5 and Section 5.1"}],"minor_comments":[{"comment":"The text says 'select an SNR of 32, equivalent to 15 dB,' but the SNR formula is in decibels by definition. Please rewrite this sentence to say that a power ratio of approximately 31.6 corresponds to 15 dB, or simply state the cutoff in dB consistently.","section":"Section 4, sentence after Eq. (5)"},{"comment":"The notation d = (fl/T)D and the instruction to use the 'first d leftmost latent dimensions' is clear for the low-frequency loss, but the high-frequency loss uses '-d' without defining negative indexing. Please define the subvector explicitly, e.g., the last d dimensions.","section":"Section 4, Eqs. (6) and (7)"},{"comment":"The text says 'We also report results on Oxford-IIIT Pet' but the main text does not show the table; it appears only in Table 2 later. Please either point to Table 2 at that location or move the relevant numbers into the main text.","section":"Section 5.2, paragraph before Table 2"},{"comment":"The radar chart is difficult to read because many methods are overlaid. Consider a table or separate plots for Known, All, and Novel, or highlight only the baselines most relevant to the comparison.","section":"Figure 6"},{"comment":"The paper says 'For FourSim, we froze the initial 11 blocks' but the appendix, Section 9.1, says 'we froze the first 9 blocks of ViT-B/16 for FourEx.' Please harmonize the numbers of frozen/fine-tuned blocks between the main text and the appendix.","section":"Section 5.1, implementation details"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid empirical core and a clear motivation, but the headline claim is stronger than the data support, particularly for DINOv2. The lack of multiple seeds is a recurring issue in this area, and I would advise the editor to require repeated runs for the main tables. The Fourier-category-information assumption is interesting but should be tested explicitly; otherwise the method risks being seen as yet another augmentation whose success is dataset-dependent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FourEx, the proposed Fourier self-supervision for fine-grained generalized category discovery, is worth a serious look. The new bit is pairing low- and high-pass Fourier-filtered views with contrastive losses on dedicated fractions of the latent dimension, and bolting that onto SelEx. On DINOv1 the empirical gains are real: +6.9 All on CUB, +3.3 on Aircraft, +2.6 on Stanford Cars, and +4.9 on Herbarium19 in the appendix. Ablations show each loss component helps, and the method is simple enough to be a plug-in.\n\nThe problem is the central claim. The abstract says Fourier Self-Supervision 'outperforms state-of-the-art methods,' but on DINOv2 the average All accuracy is exactly equal to SelEx (+0.0) and Stanford-Cars drops 2.1 All and 3.5 Novel. The paper acknowledges the Cars exception in Section 5.2, but the abstract and conclusion sweep it under the rug. That overstatement is the main reason I would not take the headline result as established.\n\nSecond, Section 4 says the appendix proves that fine details reside only in high-frequency modes. The provided appendix has no such proof. Either the proof exists and was left out, or the sentence should go. This is a load-bearing claim for the high-frequency branch, so it needs to be resolved.\n\nThird, there are no error bars or multiple-seed runs. That is common in this subfield, but it becomes more serious when one of the headline comparisons is negative. I'd want to know whether the Stanford-Cars DINOv2 drop is stable or within run-to-run noise.\n\nThe frequency-category assumption—low frequencies for broad categories, high frequencies for fine differences—is plausible and the ablations give indirect support, but it is not directly tested. That is a soft spot, not a fatal one. The SNR threshold is a quality metric, not a category-information metric; the paper does ablate SNR on CUB, which partially mitigates the concern.\n\nThe citation pattern is fine; the relevant Fourier augmentation and GCD works are covered, and the heavy overlap with the authors' own SelEx is appropriate since that is the direct baseline.\n\nBottom line: this is a solid, useful empirical paper for the category discovery subfield. It deserves a serious peer review with a request for moderate revision: reword the abstract to match the DINOv2 numbers, supply or delete the missing appendix proof, and add variance estimates if feasible. I would bring it to a reading group and I would cite it as a baseline for Fourier-based augmentation in GCD.","headline":"Solid DINOv1 results and a novel frequency-latent combination, but the abstract's SOTA claim doesn't survive DINOv2 and the promised appendix proof is missing.","tokens_in":23804,"tokens_out":4243,"would_cite":true,"duration_ms":37654,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding Fourier-filtered versions of each image as extra contrastive views improves fine-grained generalized category discovery beyond current methods, the paper reports.","keywords":["generalized category discovery","fine-grained image classification","Fourier transform","self-supervision","contrastive learning","frequency filtering","low-pass filter","high-pass filter"],"falsifier":"Take a fine-grained dataset whose class-defining differences are concentrated in mid-frequency bands, such as synthetic bird images that differ only in a mid-frequency wing pattern. If FourEx still improves over its SelEx baseline on that dataset, then the frequency-decomposition story does not explain the gains; the same test can be run by comparing the 15 dB cutoff with a data-driven cutoff chosen to maximize class-discriminative energy, and checking whether the method's advantage tracks the discriminative cutoff.","tokens_in":1464,"feed_emoji":"📡","tokens_out":1861,"duration_ms":75167,"temperature":0.7,"pith_summary":"This paper introduces Fourier Self-Supervision, a training signal for generalized category discovery (GCD) that uses the Fourier transform of each image to create two additional views: a low-pass reconstruction that keeps only coarse, abstract structure and a high-pass reconstruction that keeps fine edges and textures. The method runs contrastive learning on these filtered views in separate halves of the latent space, with the fraction of the latent dimension controlled by how much of the frequency spectrum survives at a 15 dB signal-to-noise cutoff. The authors report that adding this signal to existing GCD methods improves fine-grained discovery: on CUB-200 it raises the SelEx baseline by +6.9 in All accuracy, with further gains on FGVC-Aircraft and Stanford-Cars, and it stays ahead when the number of categories must be estimated. If correct, the insight is that category membership is stable across Fourier frequency bands, so low frequencies can provide the broad generalization needed for novel categories while high frequencies supply the fine detail needed to separate similar classes.","feed_headline":"Fourier views sharpen fine-grained category discovery","feed_subtitle":"Using low- and high-frequency image versions as extra views improves fine-grained discovery over published baselines.","key_machinery":"The central mechanism is Fourier Self-Supervision: an image is transformed to the frequency domain, a dataset-dependent cutoff is selected where the signal-to-noise ratio reaches 15 dB, and low-pass and high-pass filters produce two reconstructions used as extra positive views for contrastive learning. The ratio of the random cutoff frequency to the threshold determines what fraction of the latent dimension each band's contrastive loss is applied to, so the model is forced to represent coarse category structure in the low-frequency half and fine-grained detail in the high-frequency half. A classification term on the filtered reconstructions reinforces that low- and high-pass versions share the original image's label, and the whole set of losses is added to a base GCD objective.","core_discovery":"The paper claims that category identity is stable across the frequency spectrum of an image's Fourier transform, and that exploiting this stability as self-supervision lets a model learn both the broad attributes needed to generalize to novel categories and the fine details needed to tell similar classes apart. Concretely, each image is low-pass and high-pass filtered; the reconstructions are used as positive views in contrastive learning, with low frequencies assigned to the left half of the latent space and high frequencies to the right half. The paper shows that this dual-frequency scheme improves over the SelEx and SimGCD baselines it is attached to, and that the resulting FourEx method outperforms published state-of-the-art methods on fine-grained benchmarks in most settings, including when the number of classes is estimated rather than known.","pith_inferences":["A testable extension is to replace the fixed 15 dB SNR threshold with a per-dataset or per-class cutoff chosen by how much class-discriminative signal each frequency band carries; if the method's gains track that signal rather than a perceptual-quality threshold, the proposed mechanism is confirmed.","The same frequency-split self-supervision could be applied to other self-supervised representation learning tasks, such as domain adaptation or few-shot recognition, where low-frequency abstraction and high-frequency detail are both known to matter.","The implicit-hierarchy story suggests the low-frequency contrastive loss could serve as a lightweight hierarchical pretraining signal, potentially reducing the number of training epochs needed to learn transferable features."],"forward_implications":["Fine-grained GCD models can be improved by adding Fourier-filtered views as auxiliary positives without changing the backbone or the clustering procedure.","The gains are largest on datasets where categories differ by subtle details, such as CUB-200, FGVC-Aircraft, and Herbarium, and smaller or negative on datasets with more pronounced differences, such as Stanford-Cars.","The method remains competitive when the true number of classes is unknown and must be estimated, so the reported improvements are not an artifact of knowing the cluster count.","Because the frequency threshold is set per dataset from training images, applying the method to a new dataset requires only a forward Fourier transform pass, not new annotations.","The same loss terms can be attached to different contrastive GCD baselines, and the paper shows each attached variant improves over its own baseline."],"supporting_citations":[{"why":"sets up the generalized category discovery task, the data splits, and the base supervised and unsupervised contrastive losses that Fourier Self-Supervision extends.","marker":"[56]"},{"why":"defines the self-expertise baseline SelEx that FourEx builds on; most of the paper's reported gains are measured as improvements over this method.","marker":"[49]"},{"why":"provides the SimGCD baseline used for FourSim and serves as one of the main comparison methods for fine-grained GCD.","marker":"[60]"},{"why":"supplies the DINOv1 self-supervised ViT-B/16 pretraining used in the main experiments.","marker":"[8]"},{"why":"supplies the DINOv2 pretraining used for the second set of experiments.","marker":"[41]"},{"why":"formulates the F-principle that networks learn low-frequency components first, which motivates structuring the latent space by frequency.","marker":"[65]"},{"why":"provides the 15 dB SNR standard used to pick the dataset-dependent low- and high-pass cutoff frequencies.","marker":"[46,61]"}],"fun_headline_variants":["Frequency views sharpen fine-grained category discovery","Dual-frequency self-supervision discovers fine-grained classes","Fourier filters boost fine-grained category discovery","Seeing in low and high frequency improves category discovery","Use both ends of the frequency spectrum to find novel classes"],"cache_read_input_tokens":25856,"weakest_assumption_plain":"The load-bearing premise is that an object's category remains recognizable across its low- and high-frequency filtered versions, with low frequencies carrying broad category attributes and high frequencies carrying fine-grained distinctions; if the 15 dB cutoff is miscalibrated for a dataset, the extra contrastive losses would be noise rather than signal.","fun_headline_variants_meta":{"raw":{"variants":["Frequency views sharpen fine-grained category discovery","Dual-frequency self-supervision discovers fine-grained classes","Fourier filters boost fine-grained category discovery","Seeing in low and high frequency improves category discovery","Use both ends of the frequency spectrum to find novel classes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000253,"raw_usage":{"total_tokens":1547,"prompt_tokens":909,"completion_tokens":638,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":565}},"tokens_in":525,"tokens_out":638,"duration_ms":6162,"temperature":1.0,"reasoning_tokens":565,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:18:29.762544+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fine-grained dataset whose class-defining differences are concentrated in mid-frequency bands, such as synthetic bird images that differ only in a mid-frequency wing pattern. If FourEx still improves over its SelEx baseline on that dataset, then the frequency-decomposition story does not explain the gains; the same test can be run by comparing the 15 dB cutoff with a data-driven cutoff chosen to maximize class-discriminative energy, and checking whether the method's advantage tracks the discriminative cutoff.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022) 1, 3, 5, 9, 10, 11, 12, 20, 23, 25, 26, 27, 28","cited_arxiv_id":null,"evidence_quote":"sets up the generalized category discovery task, the data splits, and the base supervised and unsupervised contrastive losses that Fourier Self-Supervision extends."},{"cited_title":"In: ECCV (2024) 4, 9, 10, 11, 12, 20, 25, 26","cited_arxiv_id":null,"evidence_quote":"defines the self-expertise baseline SelEx that FourEx builds on; most of the paper's reported gains are measured as improvements over this method."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"provides the SimGCD baseline used for FourSim and serves as one of the main comparison methods for fine-grained GCD."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"supplies the DINOv1 self-supervised ViT-B/16 pretraining used in the main experiments."},{"cited_title":"Rastegar et al","cited_arxiv_id":null,"evidence_quote":"supplies the DINOv2 pretraining used for the second set of experiments."},{"cited_title":"In: Neural Information Processing: 26th International Conference, ICONIP 2019, Sydney, NSW, Australia, December 12–15, 2019, Proceedings, Part I 26","cited_arxiv_id":null,"evidence_quote":"formulates the F-principle that networks learn low-frequency components first, which motivates structuring the latent space by frequency."}],"review_version":1}