Pith. sign in

REVIEW 2 major objections 5 minor 17 references

SpectralGCD claims that training a classifier directly on CLIP image-concept similarities, selected by spectral filtering and aligned by forward/reverse distillation, matches or exceeds prior multimodal methods on six category-discovery ben

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 →

T0 review · deepseek-v4-flash

2026-08-02 22:13 UTC pith:DH7R3NID

load-bearing objection Strong numbers and clean ablations, but the concept dictionary likely leaks the novel class names, so the discovery claim needs a stripped-dictionary test before it convinces. the 2 major comments →

arxiv 2602.17395 v2 pith:DH7R3NID submitted 2026-02-19 cs.CV cs.AIcs.LG

SpectralGCD: Spectral Concept Selection and Cross-modal Representation Learning for Generalized Category Discovery

classification cs.CV cs.AIcs.LG
keywords generalized category discoverycross-modal representationconcept dictionaryspectral filteringknowledge distillationCLIPparametric classificationnovel class discovery
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.

Generalized Category Discovery asks a model to cluster unlabeled images into known and novel classes using only a few labels, and this paper's thesis is that the usual recipe of training a classifier on image features overfits to the known classes. SpectralGCD instead trains the classifier on CLIP's image-to-concept similarity vector, expressing each image as a mixture over a large, task-agnostic dictionary of semantic concepts. To keep this representation from drifting, the method first uses spectral filtering on softmaxed teacher similarities to retain only task-relevant concepts, then applies forward and reverse distillation to align the student's representation with the teacher's. The paper reports state-of-the-art accuracy across six benchmarks while requiring less computation than existing multimodal methods and roughly the same as unimodal ones. A sympathetic reader would take this as evidence that cross-modal concept similarities are a sufficient, cheaper representation for category discovery.

Core claim

SpectralGCD's central claim is that class identity can be read off a cross-modal representation z(x;C), the vector of cosine similarities between an image and concept names from a large agnostic dictionary, so that p(y|x) is well approximated by p(y|z(x;C)). Because the dictionary is noisy, Spectral Filtering computes the covariance of softmaxed teacher similarities over the dataset, takes its top eigenvectors, and sums squared eigenvector components weighted by eigenvalues to score and retain only concepts whose co-activation carries signal. The retained concepts define the student's representation, and forward plus reverse distillation from the same frozen teacher keeps the student's simil

What carries the argument

The central object is the cross-modal representation z_{theta,phi}(x;C), a vector of CLIP cosine similarities between an image and every concept in a task-agnostic dictionary, treated as a mixture over semantics rather than as separate visual and textual features. Two mechanisms carry the argument. Spectral Filtering forms a cross-modal covariance matrix on softmax-normalized teacher similarities, keeps the eigenspace that explains a threshold of variance, and computes a concept importance score s_j = sum_i lambda_i v_{ij}^2 to select the dictionary subset that retains class-discriminative concepts. Forward and reverse knowledge distillation losses then transfer the teacher's concept-distrib

Load-bearing premise

The fixed task-agnostic concept dictionary plus the frozen teacher's similarity structure must actually capture class identity; if the dictionary lacks the discriminating concepts, or the leading covariance eigenvectors are driven by dataset-specific background bias rather than class semantics, the representation loses the information the classifier needs.

What would settle it

Take a dataset whose classes differ along concepts missing from a chosen dictionary while sharing strong visual backgrounds; if SpectralGCD's accuracy collapses to near the unimodal baseline, or if the top selected concepts are mostly background words common to all classes, the sufficiency claim is falsified. More directly, one could test conditional independence: whether a classifier on z(x;C) matches one on x as the dictionary is varied.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A classifier trained on filtered concept similarities should keep performing on novel classes, since it is not anchored to spurious visual cues that only separate known classes.
  • Because the text encoder is frozen and teacher similarities are precomputed once, the training overhead stays close to unimodal levels even though the method is multimodal.
  • The discovery pipeline can be re-run cheaply as new unlabeled data arrive, which the paper argues is the realistic deployment setting.
  • On several benchmarks the distilled student surpasses the zero-shot accuracy of its own teacher, suggesting the method extracts more from the concept space than the teacher does at inference time.
  • Using an unfiltered dictionary hurts performance on fine-grained datasets, so concept selection is a direct accuracy lever, not just a computational one.

Where Pith is reading between the lines

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

  • If the sufficiency assumption p(y|x)=p(y|z(x;C)) holds approximately, then the bottleneck shifts from the classifier to the dictionary's concept coverage; a natural extension is to make the concept set image-adaptive rather than fixed.
  • The same spectral-filtering recipe may transfer to other label-scarce settings, such as open-set recognition or continual discovery, whenever a frozen vision-language teacher and a large concept list are available.
  • Covariance-based selection is an unsupervised proxy for class discrimination; a direct test would compare selected concepts against a supervised importance ranking and check whether the two agree on fine-grained datasets.
  • The reported robustness to generic linguistic dictionaries suggests the method could scale to 100k+ concepts with low-rank covariance approximations, trading a small accuracy dip for much broader coverage.

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

2 major / 5 minor

Summary. The paper proposes SpectralGCD, a multimodal generalized category discovery (GCD) method that replaces separate visual and textual streams with a unified cross-modal representation: CLIP image–concept cosine similarities over a large concept dictionary. A spectral filtering stage selects task-relevant concepts using the eigendecomposition of the covariance of teacher softmaxed similarities; training then uses forward and reverse distillation from a frozen teacher plus the standard SimGCD contrastive/parametric losses. Experiments on six benchmarks (CUB, Stanford Cars, FGVC-Aircraft, CIFAR-10/100, ImageNet-100) report state-of-the-art or competitive accuracy at lower training cost than previous multimodal methods, with ablations over thresholds, dictionaries, teachers, student capacity, and Old/New imbalance.

Significance. If the reported results hold, the paper makes a useful practical contribution: it shows that a parametric classifier trained on CLIP cross-modal similarities can match or exceed unimodal and prior multimodal GCD methods at a fraction of the cost, and the public code plus extensive ablations are a strength. The spectral filtering method and distillation combination are well motivated and the efficiency analysis is informative. However, the central claim that the method performs 'category discovery' using only Old class names is threatened by evidence that the concept dictionary contains exact class names of novel categories; this must be resolved before the empirical claim can be interpreted as GCD rather than class-name retrieval.

major comments (2)
  1. [§5.1, Appendix I (Table 17)] The dictionary is described as 'task-agnostic' (Abstract, §4.1) and the GCD setup assumes knowledge of Old class names only (§3). However, the default Tags dictionary is 'a list of ~22K concepts arising from several benchmark datasets' (§5.1), and the WordNet run in Table 17 selects concepts such as 'Tern', 'Warbler', 'Woodpecker', 'Vesper Sparrow', and 'Indigo Bunting' — exact CUB class names. Since z(x;C) in Eq. (4) includes similarity to these concepts, a classifier trained on z can identify novel classes by retrieving their names from the dictionary. This converts the task into semi-supervised class-name retrieval and would invalidate the claim of discovering novel categories without their names. I request a systematic check of dictionary overlap with Old/New class names for Tags, OpenImages-v7, and WordNet, and an ablation in which all Old/New class names (and near-synonyms) are rem
  2. [§4.1] The key modeling assumption p(y|x)=p(y|z(x;C)) (Section 4.1) is asserted rather than tested. It is especially important because the dictionary is finite and Table 3 shows that the choice of dictionary changes All accuracy by as much as 3–10 points (e.g., Stanford Cars, OpenImages-v7 vs Tags). If the class-name-stripped experiment from the previous comment shows a large drop, it would indicate that the representation is not sufficient in the intended sense and that the reported gains rely on explicit class-name cues. I would like to see the sufficiency assumption examined by measuring performance against oracle concepts (class names) and against a concept set with all class-name entries removed.
minor comments (5)
  1. [Eq. (4) and throughout] Notation is inconsistent: zθ,ϕ(xi; C) is defined but later written as z(x;C), and the dictionary is sometimes C and sometimes ¯C. Please unify.
  2. [Table 1] Main-table results are reported as point estimates without standard deviations; deviations appear only in Appendix Table 7. Please include them in the main table or explicitly point the reader to the appendix.
  3. [Figure 4] The figure's time comparison would be clearer if the preparation phase for each method were labeled directly on the bars (Spectral Filtering, inversion network, image-to-text assignment), rather than only in the caption.
  4. [Eq. (10)] The term 'reverse distillation' is nonstandard: L_rd = -Σ σ(z_i) log σ(z*_i) is a cross-entropy of the teacher under the student, not the reverse KL divergence. Please clarify the relationship or rename it (e.g., 'negative cross-entropy').
  5. [Appendix H] In the 80% New imbalance setting, TextGCD outperforms SpectralGCD. This is a meaningful boundary on the method's SOTA claim and deserves a sentence in the main paper, not only in the appendix.

Circularity Check

1 steps flagged

Concept dictionary appears to contain benchmark class names, making novel-class 'discovery' partially reduce to CLIP class-name retrieval; otherwise no fitted tautology.

specific steps
  1. other [Section 5.1, Eq. (4), Eqs. (8)-(9), Appendix I Table 17]
    "When not mentioned otherwise, we use the Tags concept dictionary from TextGCD as default textual information source C. ... the Tags used by TextGCD, a list of ~22K concepts arising from several benchmark datasets. ... Table 17: Top-ranked discriminative concepts ... 1 Tern 2 Warbler 3 Woodpecker 4 Vesper Sparrow 5 Indigo Bunting ..."

    Eq. (4) defines the cross-modal representation as CLIP similarities to every dictionary concept, and Eqs. (8)-(9) retain only the concepts with highest importance. The paper's own Table 17 shows that on CUB/WordNet these top concepts are ground-truth bird class names (e.g., Vesper Sparrow, Indigo Bunting). If the 'task-agnostic' dictionary is not stripped of benchmark label names, then z(x;C) already contains direct CLIP similarities to the true novel-class names. Training the classifier on W^T z and reporting New-class accuracy then reduces, at least partially, to zero-shot class-name retrieval rather than unsupervised discovery from unlabeled data. The paper reports no class-name-stripped dictionary experiment, so the central GCD claim is partly built into the input representation.

full rationale

The training and evaluation loop is otherwise not circular: accuracy is measured against ground-truth labels on external benchmarks, and the losses in Eqs. (5), (10), and (11) do not encode the ground-truth New labels. The KD Spearman result in Table 2 is an expected consequence of minimizing forward/reverse KL divergence and is reported as alignment, not as an independent prediction of accuracy. No load-bearing self-citation chain appears: citations to Mistretta et al. merely motivate cross-modal alignment and reverse KD and do not supply a uniqueness theorem or forbid alternatives. The one serious issue is dictionary leakage. Because Eq. (4) feeds every dictionary concept's CLIP similarity into the classifier, and Spectral Filtering preferentially selects class-name concepts in Table 17, some of the reported New-class accuracy is attributable to the answer key being present in the input. This is a construction-level circularity for the discovery claim, though the method still requires clustering and parametric training, so it is partial rather than total.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The method introduces no new physical or ontological entities. It rests on four domain assumptions and on thresholds selected by ablation; these are the main costs the reader pays beyond the pretrained CLIP teacher/dictionary.

free parameters (4)
  • beta_e = 0.95
    Eigenvalue cumulative-variance threshold used to set k*; chosen by ablation on Stanford Cars and CIFAR-100 (Figure 5), the same datasets used in the headline results.
  • beta_c = 0.99
    Cumulative concept-importance threshold for filtered dictionary size; chosen with beta_e in the same ablation (Figure 5).
  • lambda = 0.35
    Balance coefficient in L_cls and L_c (Eq. 5); inherited from SimGCD rather than tuned for this method.
  • CLIP logit temperature tau_t=tau_s = 0.01
    Temperature for teacher and student cross-modal similarity softmaxes (Appendix A); set equal to CLIP's default, affecting the scale of the distillation targets.
axioms (4)
  • domain assumption p(y|x)=p(y|z(x;C)) for the concept mixture z(x;C) (Section 4.1).
    The paper motivates cross-modal representations as sufficient for classification, but provides no proof that CLIP concept similarities preserve all class-relevant information; this is the foundational assumption of the method.
  • domain assumption Softmax normalization of teacher similarities amplifies class-discriminative concepts and suppresses background concepts (Section 4.2).
    The discriminativeness argument relies on CLIP's object bias and an LSA analogy; it is empirical, not proven.
  • domain assumption Dominant eigenvectors of the cross-modal covariance matrix capture task-relevant concept co-variation while small eigenvalues are noise (Section 4.2).
    Used to justify keeping only top k* components and weighting concepts by lambda_i v_i^2; no formal guarantee links this to category separability.
  • domain assumption The fixed dictionary C and frozen teacher are sufficiently broad and accurate for the task (Sections 4.1, 4.3, Limitations).
    Performance drops with weaker dictionaries/teachers (Tables 3 and 4), and the authors explicitly list this as a limitation.

pith-pipeline@v1.3.0-alltime-deepseek · 21007 in / 14623 out tokens · 140631 ms · 2026-08-02T22:13:45.589589+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of SpectralGCD: Spectral Concept Selection and Cross-modal Representation Learning for Generalized Category Discovery." pith.science (2026). https://pith.science/paper/DH7R3NID

@misc{pith2026260217395,
  author       = {Pith},
  title        = {Pith review of: SpectralGCD: Spectral Concept Selection and Cross-modal Representation Learning for Generalized Category Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DH7R3NID}},
  note         = {Machine review of arXiv:2602.17395}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Generalized Category Discovery (GCD) aims to identify novel categories in unlabeled data while leveraging a small labeled subset of known classes. Training a parametric classifier solely on image features often leads to overfitting to old classes, and recent multimodal approaches improve performance by incorporating textual information. However, they treat modalities independently and incur high computational cost. We propose SpectralGCD, an efficient and effective multimodal approach to GCD that uses CLIP cross-modal image-concept similarities as a unified cross-modal representation. Each image is expressed as a mixture over semantic concepts from a large task-agnostic dictionary, which anchors learning to explicit semantics and reduces reliance on spurious visual cues. To maintain the semantic quality of representations learned by an efficient student, we introduce Spectral Filtering which exploits a cross-modal covariance matrix over the softmaxed similarities measured by a strong teacher model to automatically retain only relevant concepts from the dictionary. Forward and reverse knowledge distillation from the same teacher ensures that the cross-modal representations of the student remain both semantically sufficient and well-aligned. Across six benchmarks, SpectralGCD delivers accuracy comparable to or significantly superior to state-of-the-art methods at a fraction of the computational cost. The code is publicly available at: https://github.com/miccunifi/SpectralGCD.

Figures

Figures reproduced from arXiv: 2602.17395 by Andrew D. Bagdanov, Lorenzo Caselli, Marco Mistretta, Simone Magistri.

Figure 1
Figure 1. Figure 1: Motivation and overview. (Left) Unimodal methods are efficient, but tend to over￾fit to spurious visual cues. (Center) Introducing textual supervision improves generalization, but increases computational overhead. (Right) SpectralGCD leverages an agnostic dictionary by se￾lecting task-relevant concepts, improving generalization while remaining as efficient as unimodal techniques. While existing multimodal … view at source ↗
Figure 2
Figure 2. Figure 2: The SpectralGCD two-phase approach. (a) Spectral Filtering uses the cross-modal covariance computed from teacher cross-modal representations to retain only its most informative components and isolate semantically relevant concepts. (b) During training, we jointly optimize the image encoder fθ, linear projection W, classifier Lψ, and MLP M using both parametric and contrastive objectives, while refining the… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of SimGCD (CLIP backbone) and Spectral￾GCD when using image features or cross-modal representations to train the classifier. Image features are slightly better on Old, while cross-modal ones improve on New. All losses operate on the cross-modal representation zθ,ϕ(xi ; C): it is projected to ui , mapped to class proba￾bilities pi , and to wi for contrastive learning. We train W and ψ, together w… view at source ↗
Figure 4
Figure 4. Figure 4: Accuracy vs. training time (s) for all methods on CUB. GET, TextGCD, and our method, require a preparation phase, whereas the uni￾modal approach SimGCD does not. By comparing multimodal and unimodal methods, we show that our approach – leveraging CLIP cross-modal representations – not only improves overall performance, but, like GET and TextGCD, also overcomes the overfit￾ting of unimodal methods to Old cl… view at source ↗
Figure 5
Figure 5. Figure 5: Ablation on thresholds βe and βc. For each threshold we report All accuracy and the resulting number of selected concepts. The black dashed line denotes performance without Spectral Filtering; the hatched bar marks the chosen configuration used for our main results. Only the unimodal SimGCD does not require a preparation phase. Our method achieves the highest accuracy while training faster than both GET an… view at source ↗
Figure 7
Figure 7. Figure 7: Performance comparison between SimGCD trained with CLIP backbone and SpectralGCD when training the classifier on ei￾ther image features (Image) or cross-modal rep￾resentations (Cross-Modal). Silhouette Score: 0.33 Image Silhouette Score: 0.53 Cross-Modal [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: t-SNE comparison across all the 20 CIFAR100 [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Total training time on CUB, Stanford Cars, CIFAR100, and ImageNet-100, including the [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] 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

17 extracted references · 4 linked inside Pith

  1. [1]

    Emergence of invariance and disentanglement in deep rep- resentations.J

    Alessandro Achille and Stefano Soatto. Emergence of invariance and disentanglement in deep rep- resentations.J. Mach. Learn. Res., 19(1):1947–1980, January

  2. [5]

    Fine-grained visual classification of aircraft.arXiv preprint arXiv:1306.5151,

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft.arXiv preprint arXiv:1306.5151,

  3. [9]

    ISBN 978-3-031-72897-6

    Springer Nature Switzerland. ISBN 978-3-031-72897-6. Grzegorz Rype ´s´c, Daniel Marczak, Sebastian Cygert, Tomasz Trzci ´nski, and Bartłomiej Twar- dowski. Category adaptation meets projected distillation in generalized continual category discov- ery. In Aleˇs Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and G¨ul Varol (eds.),Co...

  4. [10]

    Generalized category discovery via token manifold capacity learning.arXiv preprint arXiv:2505.14044, 2025a

    Luyao Tang, Kunze Huang, Chaoqi Chen, and Cheng Chen. Generalized category discovery via token manifold capacity learning.arXiv preprint arXiv:2505.14044, 2025a. Luyao Tang, Kunze Huang, Chaoqi Chen, Yuxuan Yuan, Chenxin Li, Xiaotong Tu, Xinghao Ding, and Yue Huang. Dissecting generalized category discovery: Multiplex consensus under self- deconstruction....

  5. [11]

    The caltech-ucsd birds-200-2011 dataset

    Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset

  6. [13]

    Textual knowledge matters: Cross-modality co-teaching for generalized visual class discovery

    Haiyang Zheng, Nan Pu, Wenjing Li, Nicu Sebe, and Zhun Zhong. Textual knowledge matters: Cross-modality co-teaching for generalized visual class discovery. In Aleˇs Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and G¨ul Varol (eds.),Computer Vision – ECCV 2024, pp. 41–58, Cham,

  7. [14]

    ISBN 978-3-031-72943-0

    Springer Nature Switzerland. ISBN 978-3-031-72943-0. 14 Published as a conference paper at ICLR 2026 APPENDIXA IMPLEMENTATIONDETAILS We use the CLIP with image encoder ViT-B/16 trained by OpenAI (Radford et al., 2021), to ensure fairness with the rest of the multimodal methods in GCD. We train only the last transformer block of the visual encoder. The tex...

  8. [17]

    20 Published as a conference paper at ICLR 2026 Table 17: Top-ranked discriminative concepts and common background concepts (with relative ranks) obtained using Spectral Filtering on the CUB dataset with a WordNet dictionary. Top Discriminative Concepts Common Background Concepts Rank Concept Rank Concept 1 Tern 354 Beak 2 Warbler 9,950 Wing 3 Woodpecker ...

  9. [2009]

    Benoit Dufumier, Javiera Castillo Navarro, Devis Tuia, and Jean-Philippe Thiran

    doi: 10.1109/CVPR.2009.5206848. Benoit Dufumier, Javiera Castillo Navarro, Devis Tuia, and Jean-Philippe Thiran. What to align in multimodal contrastive learning? InThe Thirteenth International Conference on Learning Representations. Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander T Toshev, and Vaishaal Shankar. Data filtering netwo...

  10. [2011]

    Get: Unlocking the multi-modal potential of clip for generalized category discovery

    Enguang Wang, Zhimao Peng, Zhengyuan Xie, Fei Yang, Xialei Liu, and Ming-Ming Cheng. Get: Unlocking the multi-modal potential of clip for generalized category discovery. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 20296– 20306, June 2025a. Guanghui Wang, Zhiyong Yang, Zitai Wang, Shi Wang, Qianqian Xu, a...

  11. [2013]

    Bagdanov

    Marco Mistretta, Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Andrew D. Bagdanov. Cross the gap: Exposing the intra-modal misalignment in CLIP via modality inversion. InThe Thirteenth International Conference on Learning Representations, 2025a. Marco Mistretta, Alberto Baldrati, Marco Bertini, and Andrew D. Bagdanov. Improving zero-shot general...

  12. [2017]

    Both contain a similar number of entries

    which spans thousands of cat- egories. Both contain a similar number of entries. Here, we report TextGCD results using only these dictionaries, excluding the auxiliary LLM-generatedAttributesdictionary, to ensure that both meth- ods operate under the same textual constraints. Results on CUB are consistent with those observed on the other datasets:Tagsprov...

  13. [2020]

    Concept bottleneck models

    11 Published as a conference paper at ICLR 2026 Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In Hal Daum´e III and Aarti Singh (eds.),Proceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pp. 5338–5348....

  14. [2021]

    Learn to categorize or categorize to learn? self- coding for generalized category discovery

    12 Published as a conference paper at ICLR 2026 Sarah Rastegar, Hazel Doughty, and Cees Snoek. Learn to categorize or categorize to learn? self- coding for generalized category discovery. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (eds.),Advances in Neural Information Processing Systems, volume 36, pp. 72794–72818. Curran Assoc...

  15. [2023]

    Asano, Hazel Doughty, and Cees G

    Sarah Rastegar, Mohammadreza Salehi, Yuki M. Asano, Hazel Doughty, and Cees G. M. Snoek. Selex: Self-expertise in fine-grained generalized category discovery. In Ale ˇs Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and G ¨ul Varol (eds.),Computer Vision – ECCV 2024, pp. 440–458, Cham,

  16. [2025]

    Automatically discovering and learning new visual categories with ranking statistics.arXiv preprint arXiv:2002.05714,

    Kai Han, Sylvestre-Alvise Rebuffi, Sebastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Automatically discovering and learning new visual categories with ranking statistics.arXiv preprint arXiv:2002.05714,

  17. [4090]

    In Table 8 we include additional dataset details, regarding the samples count and splits used in our evaluation

    Reported performance is averaged on three different seeds (see Table 7 for standard deviations). In Table 8 we include additional dataset details, regarding the samples count and splits used in our evaluation. APPENDIXB BALANCE BETWEENOLD ANDNEWPERFORMANCE INGCD In this section, we discuss how unimodal approaches suffer from theOld/Newtradeoff more severe...