acceptodds
Under review as a conference paper at ICLR 2027

SPARC: Learning Spectrally Stable Assignments for Generalized Category Discovery

Abstract

Generalized Category Discovery (GCD) aims to recognize known categories and discover novel categories in unlabeled data. Despite recent progress driven by strong representation learning, GCD models may rely on appearance cues such as texture, illumination, and background. These cues can distort feature similarity and lead to unstable category assignments. Existing methods mainly enforce consistency between spatially augmented views, while the effect of frequency-domain appearance changes on category assignments remains underexplored. Moreover, prior Fourier-based methods typically use frequency operations for augmentation or feature enhancement, rather than explicitly improving assignment stability. To address this issue, we propose \name, a plug-and-play module for SPectral Assignment Regularization for Categories. SPARC contains two components. Image-level Consistency Regularization (ICR) mixes low-frequency amplitudes from different images while retaining the source phase, and encourages the mixed image to preserve the assignment of the clean image. Token-level Consistency Regularization (TCR) identifies patch tokens that remain stable after spectral filtering, pools them into a Stable Semantic Token (SST), and aligns its assignment with that of the global descriptor. For parametric hosts, SST is additionally used as the readout during inference. Experiments on representative GCD baselines with DINO and DINOv2 demonstrate performance improvements across CIFAR-10/100, ImageNet-100, CUB, Stanford-Cars, and FGVC-Aircraft.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Generalized Category Discovery, Fourier Transform

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_uzguOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SPARC, a plug-and-play framework for generalized category discovery that improves assignment stability through spectral regularization, combining image-level consistency, token-level consistency, and stable semantic token selection.

Critical strengths
  1. The paper explores spectral assignment consistency in GCD, providing a meaningful perspective on reducing sensitivity to appearance variations.
  2. Extensive experiments across multiple datasets and backbones demonstrate consistent average improvements on fine-grained benchmarks.
Critical weaknesses
  1. Although the paper emphasizes spectral consistency as its central contribution, Table 3 shows that SST alone provides the largest individual performance improvement. This raises concerns about whether the main gains originate from spectral regularization or simply from replacing the CLS token with a more discriminative local representation.
  2. Table 3 suggests complementary effects among ICR, SST, and TCR, but notably omits the ICR+TCR configuration without SST readout replacement. This missing combination prevents a complete assessment of the interaction between the two proposed spectral regularizers and makes it difficult to isolate their contributions from SST.
  3. Several related modular GCD approaches have emerged during 2025–2026. Since SPARC is explicitly presented as a plug-and-play framework, comparisons should evaluate these methods under the same host architecture and experimental settings, rather than relying primarily on standalone performance comparisons.
  4. Although SPARC improves its host baselines on average, its absolute performance on coarse-grained benchmarks remains below several existing approaches, with some configurations showing negligible or negative improvements. The paper lacks a convincing explanation of why the proposed spectral regularization is substantially more effective on fine-grained datasets and whether its benefits are fundamentally task-dependent.
Questions for authors

please see weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer uzgu (1/2)

[Q1] Ablation

Thank you for your constructive reivew. Your concern relates to the effectiveness of two distinct pieces of SPARC: (i) the Stable Semantic Token (SST) and (ii) spectral assignment regularizers (ICR/TCR). Firstly, we would like to emphasize that SST itself is a core contribution, not a side effect to be explained away. It provides a frequency-domain route to more discriminative local cues. To our knowledge, this form of spectral token selection for GCD has not appeared in prior work. Secondly, from Tab. 3 we can conclude that the spectral regularizers still help beyond the readout, even though SST alone already provide a huge improvement. Holding the readout fixed, adding ICR+TCR lifts SST from 61.1 to 64.7 Avg All (CUB 63.3→68.0, Stanford-Cars 63.0→67.6, FGVC-Aircraft 57.1→58.5). Thirdly, we further conduct an ablation on the SelEx host. In this scenario, SST is not used as a readout, which allows us to clearly figure out the contribution of spectral regularizers. In this table, by comparing the second and third row with the baseline (first row), we can see that each one regularizer alone yields an improvemen. Besides, Avg All rises from 63.1 to 70.2 (CUB 73.6→79.6, Stanford-Cars 58.5→68.6, FGVC-Aircraft 57.1→62.4) with two regularizers working together.

Table. Component ablation (SelEx host, DINO)

ICR/TCR ablation on SelEx (DINO); SST is not applied as a CLS replacement. Metrics are All / Old / New ACC (%).

ICRTCRCUBStanford-CarsFGVC-AircraftAverage
✗✗73.6 / 75.3 / 72.858.5 / 75.6 / 50.357.1 / 64.7 / 53.363.1 / 71.9 / 58.8
✓✗75.9 / 76.6 / 75.663.4 / 75.1 / 57.761.2 / 65.0 / 59.366.8 / 72.2 / 64.2
✗✓77.7 / 77.0 / 78.066.4 / 76.4 / 61.662.1 / 67.1 / 59.668.7 / 73.5 / 66.4
✓✓79.6 / 76.5 / 81.168.6 / 76.1 / 65.062.4 / 67.1 / 60.170.2 / 73.2 / 68.7

[Q2] Ablation

Thank you for pointing this out. We would like to showcase the full component grid in the SimGCD host ablation table below, including the ICR+TCR configurtaion without SST. With the complete 3 components ablation table, we observe that the qualitative picture is consistent with the paper, that is, three components are complementary and generally most effective when used jointly.

Additionlly, since SST is not as a readout for SelEx, we provide a 2 components ablation on SelEx as above, which perfectly isolates their contribution from SST. This table shows that ICR and TCR both help improve accuracy for SelEx and that they works better together (Average All moves from 63.1 to 66.8 with ICR, to 68.7 with TCR, and to 70.2 with both).

Table. Component ablation (SimGCD host, DINO)

Component ablation of ICR, SST, and TCR on CUB, Stanford-Cars, and FGVC-Aircraft. Metrics are All / Old / New ACC (%).

ICRSSTTCRCUBStanford-CarsFGVC-AircraftAverage
✗✗✗60.3 / 65.6 / 57.753.8 / 71.9 / 45.054.2 / 59.1 / 51.856.1 / 65.5 / 51.5
✓✗✗62.0 / 66.7 / 61.455.8 / 72.4 / 47.856.5 / 63.4 / 53.158.1 / 67.5 / 54.1
✗✓✗63.3 / 70.5 / 59.763.0 / 80.6 / 54.557.1 / 64.8 / 53.361.1 / 72.0 / 55.8
✗✗✓63.5 / 67.7 / 61.454.0 / 71.9 / 45.455.5 / 61.8 / 52.357.7 / 67.1 / 53.0
✓✓✗63.0 / 71.9 / 58.663.1 / 79.4 / 55.356.4 / 61.9 / 53.760.8 / 71.1 / 55.9
✓✗✓63.5 / 68.8 / 60.859.0 / 78.0 / 49.856.8 / 63.4 / 53.559.8 / 70.1 / 54.7
✗✓✓66.5 / 70.7 / 59.863.9 / 79.8 / 56.259.2 / 65.7 / 55.962.2 / 72.1 / 57.3
✓✓✓68.0 / 71.7 / 66.167.6 / 81.2 / 61.058.5 / 66.2 / 54.764.7 / 73.0 / 60.6

AuthorsOpenReview ↗

Response to Reviewer uzgu (2/2)

[Q3] Plug-and-play

Thank you for your comments. Our earlier emphasis on standalone absolute numbers was mainly to show that SPARC can push the accuracy boundary of fine-grained GCD; for a plug-and-play claim, a same-host protocol is the fairer yardstick, and we adopt it here. We report modular add-ons on fixed SimGCD and SelEx hosts (DINO / DINOv2), with host-relative Δ\Delta, against recent modular methods (e.g., APL, ConGCD, LEAH, CPF-GCD). SimGCD and SelEx represent parametric versus non-parametric pipelines.

From this table, we can observe that our SPARC consistently improves over both SimGCD and SelEx hosts under DINO and DINOv2. For example, on DINO + SimGCD, SPARC raises the average All ACC from 56.1 to 64.7, outperforming APL (60.4), ConGCD (57.1), LEAH (64.4), and CPF-GCD (59.1). Similar gains are observed on SelEx and DINOv2, showing that SPARC is a general and effective plug-and-play module.

Table. Performace on different hosts and backbones

BackboneMethodVenueCUBStanford-CarsFGVC-AircraftAverage
DINOSimGCDICCV 202360.3 / 65.6 / 57.753.8 / 71.9 / 45.054.2 / 59.1 / 51.856.1 / 65.5 / 51.5
DINO+ APLCVPR 202564.5 / 68.1 / 62.160.1 / 77.6 / 51.256.6 / 60.2 / 54.860.4 / 68.6 / 56.0
DINO+ ConGCDICCV 202561.6 / 65.1 / 59.554.5 / 72.2 / 47.855.0 / 58.4 / 53.157.1 / 65.2 / 53.5
DINO+ LEAHAAAI 202669.4 / 76.3 / 65.965.5 / 80.4 / 58.258.3 / 64.7 / 55.164.4 / 73.8 / 59.7
DINO+ CPF-GCDICML 202663.9 / 65.5 / 63.157.6 / 73.7 / 49.855.8 / 59.1 / 54.259.1 / 66.1 / 55.7
DINO+ SPARC–68.0 / 71.7 / 66.167.6 / 81.2 / 61.058.5 / 66.2 / 54.764.7 / 73.0 / 60.6
Inc.+7.7 / +6.1 / +8.4+13.8 / +9.3 / +16.0+4.3 / +7.1 / +2.9+8.6 / +7.5 / +9.1
DINOSelExECCV 202473.6 / 75.3 / 72.858.5 / 75.6 / 50.357.1 / 64.7 / 53.363.1 / 71.9 / 58.8
DINO+ ConGCDICCV 202581.7 / 80.4 / 82.457.5 / 77.5 / 47.962.5 / 70.2 / 58.767.3 / 76.0 / 63.0
DINO+ CPF-GCDICML 202679.8 / 80.5 / 79.458.9 / 79.0 / 49.261.8 / 69.0 / 58.166.8 / 76.2 / 62.2
DINO+ SPARC–79.6 / 76.5 / 81.168.6 / 76.1 / 65.062.4 / 67.1 / 60.170.2 / 73.2 / 68.7
Inc.+6.0 / +1.2 / +8.3+10.1 / +0.5 / +14.7+5.3 / +2.4 / +6.8+7.1 / +1.3 / +9.9
DINOv2SimGCDICCV 202371.5 / 78.1 / 68.371.5 / 81.9 / 66.663.9 / 69.9 / 60.969.0 / 76.6 / 65.3
DINOv2+ APLCVPR 202575.1 / 79.1 / 73.273.4 / 87.6 / 66.768.8 / 74.1 / 66.672.4 / 80.3 / 68.8
DINOv2+ SPARC–80.3 / 82.8 / 79.079.8 / 90.5 / 74.773.9 / 74.9 / 73.378.0 / 82.7 / 75.7
Inc.+8.8 / +4.7 / +10.7+8.3 / +8.6 / +8.1+10.0 / +5.0 / +12.4+9.0 / +6.1 / +10.4
DINOv2SelExECCV 202487.4 / 85.1 / 88.582.2 / 93.7 / 76.779.8 / 82.3 / 78.683.1 / 87.0 / 81.3
DINOv2+ ConGCDICCV 202586.3 / 87.4 / 85.879.8 / 93.1 / 73.381.7 / 83.3 / 81.082.6 / 87.9 / 80.0
DINOv2+ SPARC–90.5 / 88.6 / 91.584.0 / 93.5 / 79.781.6 / 83.1 / 80.885.4 / 88.4 / 84.0
Inc.+3.1 / +3.5 / +3.0+1.8 / -0.2 / +3.0+1.8 / +0.8 / +2.2+2.3 / +1.4 / +2.7

[Q4] Performance on coarse-grained benchmarks

Thank you for your insightful comments. Indeed, the magnitude of the gain is task-dependent, but our SPARC improves the host on both fine- and coarse-grained benchmarks on average. From the rightmost 3 columns of both Tab.1 and Tab.2, we can observe consistent improvements of our methods across different datasets, hosts and backbones. The reason that SPARC does better on fine-grained datasets is likely that fine-grained categories often differ in local parts and textures; SST + ICR/TCR directly target this local-cue bottleneck, which leads to larger gains. However, on coarse-grained data, classes are already well separated by global semantics, so there is little headroom left. Additional local spectral constraints therefore bring only modest margins and can occasionally over-regularize. Nevertheless, we still observe notable improvements - for example, + 4.0 in All and + 4.8 in New on SelEx + SPARC with DINO as well as + 2.8 in All and +3.7 in New on SimGCD + SPARC with DINOv2 for ImageNet-100 - and the overall average improvement remains positive.

We hope this addresses your concerns and welcome any follow-up questions.

Reviewer_QajuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SPARC adds spectral consistency losses to a GCD model. It mixes image amplitudes and pools patch tokens that remain stable under filtering, then regularizes their category assignments. For parametric hosts, the pooled Stable Semantic Token also replaces CLS at inference. The main experiments use SimGCD and SelEx on six datasets, with several pretrained backbones.

Critical strengths
  1. The fine-grained gains are convincing. With DINO, mean All accuracy goes from 56.1% to 64.7% on SimGCD and from 63.1% to 70.2% on SelEx. Seeing improvements on both hosts, followed by DINOv2, CLIP, and DINOv3 results, makes the module worth considering.
  2. Table 3 separates the readout change from the two losses. The assignment-stability measurements in Appendix C are also helpful: they test the behavior the method is intended to improve, beyond showing selected attention maps.
Critical weaknesses
  1. It is still difficult to tell how much specifically requires spectral assignment regularization. SST alone gives a large Cars gain, from 53.8% to 63.0%. There is no matched comparison against uniform or attention-based pooling, nor feature consistency under the same spectral perturbations. The combination works, but the stronger explanation about why it works needs these distinctions.
  2. Some basic evaluation details are missing. ICR adds a backbone view, and SST still needs spectral selection at inference, yet no timing is given. The tables also lack seed variability. Appendix A.3 calls the defaults “CUB-validated” without explaining the validation split or which labels were available; this needs to be explicit in a category-discovery experiment.
Questions for authors
  1. Have you compared with matched pooling or feature-consistency controls? Existing results would help separate the spectral contribution from the readout change.
  2. How were the CUB defaults selected? Please specify label access and report any available seed variability and training/inference timings.
7 comments
AuthorsOpenReview ↗

Response to Reviewer Qaju (1/2)

[Q1] and [Q3] Comparisons against pooling methods and feature consistency

Thank you for your constructive question. We would like to provide three more experiments to clarify the contribution of our components: a comparison with standard pooling methods, a comparison of different consistency objectives and an ablation on the SelEx host.

First, we compare SST with mean and max pooling. Since various pooling methods exist, we pick the two most common types - uniform mean and max pooling - alongside our spectral SST. On the three-dataset average, All ACC increases from 56.1 with the CLS baseline to 60.5 with mean pooling, 55.1 with max pooling and to 61.1 with our SST. ICR+TCR then adds 2.2 Avg All on mean pooling, and 3.6 on SST. Max pooling + ICR + TCR does not work well because the quality of selected tokens are not good enough, which also pollutes TCR. This suggests that the quality of the selected tokens matters. Overall, our SST provides a stronger readout and ICR/TCR further improves it.

Table. Pooling methods ablation (SimGCD host, DINO).

Matched comparison of mean / max pooling and spectral SST under the same SimGCD host, with and without ICR+TCR. Metrics are All / Old / New.

ICRTCRMeanMaxSSTCUBStanford-CarsFGVC-AircraftAverage
✗✗✗✗✗60.3 / 65.6 / 57.753.8 / 71.9 / 45.054.2 / 59.1 / 51.856.1 / 65.5 / 51.5
✗✗✓✗✗62.7 / 70.2 / 58.962.9 / 78.1 / 55.656.0 / 59.1 / 54.460.5 / 69.1 / 56.3
✗✗✗✓✗58.9 / 68.1 / 54.353.6 / 72.6 / 44.552.9 / 60.1 / 49.255.1 / 66.9 / 49.3
✗✗✗✗✓63.3 / 70.5 / 59.763.0 / 80.6 / 54.557.1 / 64.8 / 53.361.1 / 72.0 / 55.8
✓✓✗✗✗63.5 / 68.8 / 60.859.0 / 78.0 / 49.856.8 / 63.4 / 53.559.8 / 70.1 / 54.7
✓✓✓✗✗65.1 / 68.5 / 63.464.6 / 81.6 / 56.458.4 / 64.2 / 55.662.7 / 71.4 / 58.5
✓✓✗✓✗62.1 / 71.2 / 57.650.6 / 67.1 / 42.749.7 / 57.9 / 45.554.1 / 65.4 / 48.6
✓✓✗✗✓68.0 / 71.7 / 66.167.6 / 81.2 / 61.058.5 / 66.2 / 54.764.7 / 73.0 / 60.6

Second, we compare consistency objective while keeping the spectral perturbation fixed. Applying feature-cosine loss to the backbone representation raises average All from 56.1 to 60.5, while our SPARC reaches 64.7. Hence, performing consistency learning in the assignment space is more effective than performing it directly in the feature space, and is therefore preferable.

Table. Comparison of different assignment methods (SimGCD host, DINO).

Feature cosine uses the same spectral perturbation, with the loss on the backbone representation. Metrics are All / Old / New.

Assignment methodCUBStanford-CarsFGVC-AircraftAverage
SimGCD60.3 / 65.6 / 57.753.8 / 71.9 / 45.054.2 / 59.1 / 51.856.1 / 65.5 / 51.5
Feature cosine61.9 / 71.8 / 56.964.2 / 78.9 / 57.155.5 / 63.6 / 51.560.5 / 71.4 / 55.2
SPARC68.0 / 71.7 / 66.167.6 / 81.2 / 61.058.5 / 66.2 / 54.764.7 / 73.0 / 60.6

To seperate the spectral contribution from the readout change, we also run a component ablation on SelEx host. This host does not use SST as its readout: the clustering feature remains CLS. ICR alone and TCR alone each raise All on every dataset. Average All moves from 63.1 to 66.8 with ICR, to 68.7 with TCR, and to 70.2 with both. The gain sits mainly on New classes, where the average rises from 58.8 to 68.7. Consequently, we can conclude that ICR and TCR contribute to our model even without SST.

Table. Component ablation (SelEx host, DINO).

SST is not used as a replacement for CLS. Metrics are All / Old / New.

ICRTCRCUBStanford-CarsFGVC-AircraftAverage
✗✗73.6 / 75.3 / 72.858.5 / 75.6 / 50.357.1 / 64.7 / 53.363.1 / 71.9 / 58.8
✓✗75.9 / 76.6 / 75.663.4 / 75.1 / 57.761.2 / 65.0 / 59.366.8 / 72.2 / 64.2
✗✓77.7 / 77.0 / 78.066.4 / 76.4 / 61.662.1 / 67.1 / 59.668.7 / 73.5 / 66.4
✓✓79.6 / 76.5 / 81.168.6 / 76.1 / 65.062.4 / 67.1 / 60.170.2 / 73.2 / 68.7

AuthorsOpenReview ↗

Response to Reviewer Qaju (2/2)

[Q2] and [Q4] Timing, seed variability and CUB defaults

Thank you for your comments. We address each point below.

For the timing, we compare the SimGCD host with the full SPARC model. On the controlled SimGCD and SimGCD + SPARC runs, training rises from about 28 s to 41 s per epoch and evaluation from about 14 s to 20 s, which is acceptable. On the full datasets the epoch time rises by about 23 s on CUB, 13 s on Stanford-Cars, and 16 s on FGVC-Aircraft with only around +15% increase for Stanford-Cars and FGVC-Aircraft.

Table. Wall-clock.

The first two rows are the controlled measurements. The dataset rows are full-epoch time.

MeasurementBaseline+ SPARCΔ
Train (s/epoch)~28~41+13
Eval (s/epoch)~14~20+6
CUB, epoch~37 s~60 s+23 s
Stanford-Cars, epoch~82 s~94 s+13 s
FGVC-Aircraft, epoch~109 s~125 s+16 s

For seed variability, we repeat full SPARC with differnt seeds. All accuracy is 67.1 ± 0.2 on CUB, 66.4 ± 1.0 on Stanford-Cars, and 58.4 ± 0.3 on FGVC-Aircraft. Every seed stays substantially above the published SimGCD All numbers, 60.3, 53.8, and 54.2.

Table. Three seeds of SPARC on DINO.

Entries are All / Old / New. The mean and standard deviation are across seeds.

MethodCUBStanford-CarsFGVC-AircraftAverage
SimGCD60.3 / 65.6 / 57.753.8 / 71.9 / 45.054.2 / 59.1 / 51.856.1 / 65.5 / 51.5
SPARC, seed 167.2 / 69.7 / 66.065.5 / 79.4 / 58.858.3 / 66.3 / 54.463.7 / 71.8 / 59.7
SPARC, seed 267.2 / 68.4 / 66.767.4 / 80.3 / 61.258.1 / 62.6 / 55.764.2 / 70.4 / 61.2
SPARC, seed 366.9 / 70.5 / 65.166.2 / 79.9 / 59.658.7 / 65.0 / 55.663.9 / 71.8 / 60.1
mean ± std67.1 ± 0.2 / 69.5 ± 1.1 / 65.9 ± 0.966.4 ± 1.0 / 79.9 ± 0.5 / 59.9 ± 1.258.4 ± 0.3 / 64.6 ± 1.9 / 55.2 ± 0.763.9 ± 0.3 / 71.3 ± 0.8 / 60.3 ± 0.8

CUB defaults are selected following previous efforts, including GCD (Vaze et al., 2022) and DebGCD (Liu, 2025), on a standout validation set, which is the test split of the dataset. Only known classes labels are used in the procedure and the best parameters are choosen according to the best performance on these classes. For hyperparameter performance comparison, we also apply a staged grid search of the parameters include ρ\rho, κ\kappa, TopK, λimg\lambda_{\mathrm{img}}, and λtok\lambda_{\mathrm{tok}}, which can be found in main paper's Tab. 4, Fig. 4, Fig. 5, and Appendix Tab. 7–9.

We hope this addresses your concerns, and we welcome any follow-up questions.

Thank you for the pooling controls, SelEx ablation, and timing and seed results. These strengthen the empirical evidence. Two points still need clarification: are the CUB samples used for hyperparameter selection also included in the final evaluation, and does the feature-consistency comparison use the same readout and other training settings? Pending these clarifications, I maintain my rating of 2.

Reviewer_CYsSOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Based on spectral perspective, this paper propose the plug-and-play module SPARC in GCD problem. It utilizes ICR and TCR to regularize the low frequency interaction from other photos and enhance the concentration on spectrally stable patches by TopK voted SST. Experiments show improvements in assignment stability and clustering accuracy across multiple GCD benchmarks.

Critical strengths
  1. The paper is well-written and technically sound.
  2. The motivation is clear.
Critical weaknesses
  1. The notation in Figure1 (a) makes me confused. There are six slots predicted slots and no SimGCD metioned in (a) subfig.
  2. In ICR, how many / how to choose the target images to mixture in the Fourier domain?
  3. In SimGCD, the tempreture for teacher is smaller than student, aiming to give a more confident(sharp) distribution. What's the strategy here for temperture set in ICR and TCR?
  4. Why does stability under low-pass filtering indicate semantic reliability? Smooth background regions may also remain stable, while discriminative details may be altered.
  5. In ablation, there exists a suprising gap for ICR+SST+TCR compared to the others, why it happen? And lacks the ablation of ICR & TCR without SST to readout.
Questions for authors

See above.

7 comments
AuthorsOpenReview ↗

Response to Reviewer CYsS (1/2)

[Q1] Notation in Fig. 1(a)

Thank you for your careful questions. The host illustrated in Fig. 1(a) is exactly SimGCD; we will state this explicitly in both the panel label and the caption. We will also correct the "Top-5" typo to "Top six" in the caption.

[Q2] ICR

Thanks for the comments. We are happy to clarify our ICR paring machanism. Each source image is paired with exactly one target image. The target is obtained by cyclically shifting the current mini-batch, which guarantees that no sample is paired with itself. The mixing weight λ\lambda is drawn independently from Unif[0,1]\mathrm{Unif}[0,1]; we mix the Fourier amplitudes while keeping the source phase. We do not mine hard targets, as a random in-batch partner is sufficient for the consistency regularizer.

[Q3] Temperature strategy

Thanks for the comments. In fact, ICR and TCR reuse the host SimGCD temperatures and do not introduce a separate schedule. The teacher (stop-gradient side) is warmed from 0.07 to 0.04 over the first 30 epochs, matching SimGCD’s clustering teacher. The student side optimized by ICR/TCR stays at a fixed temperature of 0.1, matching SimGCD’s student temperature. Spectral consistency is therefore enforced in the same confidence geometry as the host distillation loss, so ICR/TCR do not fight a differently calibrated assignment space.

[Q4] Low-pass stability

Thanks for your insightful comments. Firstly, our SST is selected not only based on a low-pass filter but on a composition of Gaussian spectral mask, low-pass probe and TopK selection/pooling. The link to semantics lies in the selection, not merely in the blur: distinctive object parts typically collect consistent votes across channels and enter the pooled descriptor, whereas large flat backgrounds—despite being stable—rarely dominate the TopK set. Secondly, we agree that some tokens of smooth background regions may also remain stable and being chosen. However, we do not claim our SST as a pure foreground selector, as contextual background can itself be class-informative in fine-grained recognition. The criterion is not “find foreground,” but “find a stable, high-consensus local quantity that is useful for discrimination.” Thirdly, this design is empirically effective. Fig. 5 shows that selected tokens concentrate on distinctive parts. Matched SimGCD ablation in Tab.3 also indicates that SST alone already improves over the CLS host (Avg All: CUB 60.3→63.3, Stanford-Cars 53.8→63.0, FGVC-Aircraft 54.2→57.1).

AuthorsOpenReview ↗

Response to Reviewer CYsS (2/2)

[Q5] Ablation

We appreciate this question and would be glad to provide the complete ablation table including ICR + TCR, as presented below. From table below, the qualitative picture is consistent with the paper. On SimGCD, every single factor improves over the host on average. SST is the strongest individual add-on, which offers a good anchor for two regularizers. Pairwise combinations generally improve further and three components together usually achieve best results. It shows that three components are complementary - SST supplies a stable local readout; ICR/TCR regularize assignments under spectral perturbations.

Table. Component ablation (SimGCD host, DINO)

Component ablation of ICR, SST, and TCR on CUB, Stanford-Cars, and FGVC-Aircraft. Metrics are All / Old / New ACC (%).

ICRSSTTCRCUBStanford-CarsFGVC-AircraftAverage
✗✗✗60.3 / 65.6 / 57.753.8 / 71.9 / 45.054.2 / 59.1 / 51.856.1 / 65.5 / 51.5
✓✗✗62.0 / 66.7 / 61.455.8 / 72.4 / 47.856.5 / 63.4 / 53.158.1 / 67.5 / 54.1
✗✓✗63.3 / 70.5 / 59.763.0 / 80.6 / 54.557.1 / 64.8 / 53.361.1 / 72.0 / 55.8
✗✗✓63.5 / 67.7 / 61.454.0 / 71.9 / 45.455.5 / 61.8 / 52.357.7 / 67.1 / 53.0
✓✓✗63.0 / 71.9 / 58.663.1 / 79.4 / 55.356.4 / 61.9 / 53.760.8 / 71.1 / 55.9
✓✗✓63.5 / 68.8 / 60.859.0 / 78.0 / 49.856.8 / 63.4 / 53.559.8 / 70.1 / 54.7
✗✓✓66.5 / 70.7 / 59.863.9 / 79.8 / 56.259.2 / 65.7 / 55.962.2 / 72.1 / 57.3
✓✓✓68.0 / 71.7 / 66.167.6 / 81.2 / 61.058.5 / 66.2 / 54.764.7 / 73.0 / 60.6

Additionally, to illustrate the effectiveness of ICR and TCR without SST, we conduct an ablation study on SelEx host. In this setting, where readout remains CLS, ICR and TCR each improve the host, and the combination yields best on average.

Table. Component ablation (SelEx host, DINO)

ICR/TCR ablation on SelEx (DINO); SST is not applied as a CLS replacement. Metrics are All / Old / New ACC (%).

ICRTCRCUBStanford-CarsFGVC-AircraftAverage
✗✗73.6 / 75.3 / 72.858.5 / 75.6 / 50.357.1 / 64.7 / 53.363.1 / 71.9 / 58.8
✓✗75.9 / 76.6 / 75.663.4 / 75.1 / 57.761.2 / 65.0 / 59.366.8 / 72.2 / 64.2
✗✓77.7 / 77.0 / 78.066.4 / 76.4 / 61.662.1 / 67.1 / 59.668.7 / 73.5 / 66.4
✓✓79.6 / 76.5 / 81.168.6 / 76.1 / 65.062.4 / 67.1 / 60.170.2 / 73.2 / 68.7

We hope this addresses your concerns and welcome any follow-up questions.

Thank you for your response.

  • I recommend you to fix and highlight the change of Q1, Q3 in the article.
  • For Q2, the question still exists. ICR disturbs the low-pass energy in Fourier domain as a special Augmentation process, but is different to the traditional Augmentation method in SSL only use aug-data on it own. Author should clearly clarify the procedure in the article, and give a more extreme condition that when disturbed the low-pass energy not by mini-batch but by a random low-pass noise.
  • For Q5, author gives the full ablation result but not well explained the improvement of it. This makes me still classify this article method as a technical one with some concern.

Therefore, I still remain my confidence and rate.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions

We have revised the paper and would like to invite the reviewers to take a look. Following the reviewers' suggestions, we have made the following major updates:

Section 1: Changed the typo "Top-5" to "Top six"; and add "assignment space of SimGCD" in the teaser figure.

Section 4: Added one line of "ICR + TCR" in ablation table.

Section 4.1: Change "A donor x′x'" to "A target image, x′x' ".

Appendix B.2: Added ablation study on SelEx host.

Appendix B.3: Added experiments with different pooling methods.

Appendix B.4: Added experiments with different space for consistency learning.

Appendix B.8: Added experiments with different amplitude mixing strategy.

Appendix B.9: Added experiments comparing feature discriminability between generic and fine-grained datasets.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.