Pith. sign in

REVIEW 4 major objections 7 minor 44 references

FaceGCD: Generalized Face Discovery via Dynamic Prefix Generation

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A hypernetwork-generated per-image prefix lets a frozen ViT discover new face identities while recognizing known ones, beating GCD methods and ArcFace on six face benchmarks.

desk verdict New GFD benchmark plus a hypernetwork-based dynamic prefix method; gains are consistent, but unexamined identity overlap between pretraining and evaluation undercuts the absolute open-world discovery claim. read the letter →

arxiv 2507.22353 v1 pith:G4PSOQ3Z submitted 2025-07-30 cs.CV

Yunseok Oh , Dong-Wan Choi This is my paper · ORCID
classification cs.CV
keywords generalizedfacediscoveryidentificationnovelidentityprefixtuninghypernetworkVisionTransformerclusteringopen-worldrecognition
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

The paper introduces Generalized Face Discovery (GFD), a task that requires a face recognition model to classify known identities and simultaneously group previously unseen faces into new identities. It claims that generic GCD methods fail on this task because faces form hundreds of visually similar, fine-grained classes, and that this failure can be overcome by making the feature extractor instance-specific. Its proposed FaceGCD generates, for each input image, lightweight layer-wise prefix tokens via a hypernetwork, modulating a frozen ViT backbone at every layer. The authors report that FaceGCD outperforms all tested GCD baselines and the ArcFace face-recognition baseline on six benchmarks built from YouTube Faces and CASIA-WebFaces, with 82.7% overall accuracy on the YTF 1000 set. If correct, this would establish a new state of the art for open-world face recognition that unifies identification and discovery.

What carries the argument

The load-bearing mechanism is the hypernetwork-conditioned prefix generator. A two-layer MLP hypernetwork takes global-average-pooled layer-wise features from a frozen ViT-B/16 and produces per-layer weight and bias parameters for a down–up projection block; that block transforms randomly initialized key and value prefix tokens into layer-wise prefixes, which are prepended to the self-attention keys and values of every transformer layer. Because the hypernetwork's conditioning input is the image's own features, the resulting feature extractor changes per instance rather than being shared. The generative procedure is what the ablation study isolates: replacing it with a static prefix pool drops overall accuracy by 40.3 points, and replacing it with a large static prefix generator (over 10x more parameters) still lags by 6.0 points.

What would settle it

Compute the identity overlap between MS1MV3, the pretraining set of the ViT backbones, and the identities in YTF/CASIA used for evaluation. If a substantial fraction of the 'novel' test IDs also appear in MS1MV3, the reported novel-ID accuracy overstates genuine discovery. A complementary check trains FaceGCD with a ViT pretrained only on non-face data; if the advantage over static baselines then collapses, the dynamic prefix mechanism is likely benefiting from memorized face features rather than from its own conditioning.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that instance-adaptive feature extraction, obtained by injecting prefix tokens whose weights are produced on the fly by a hypernetwork, is what makes generalized face discovery work. Existing GCD methods, designed for heterogeneous datasets like CIFAR-100, produce features that do not separate fine-grained face identities; ArcFace, though a strong closed-set recognizer, does not separate novel identities. FaceGCD instead conditions a frozen ViT on each input image: the hypernetwork reads layer-wise features from the frozen backbone and emits the parameters of a small prefix generator, which turns random prefix tokens into key/value prefixes for every self-attention layer. With the same semi-contrastive loss and semi-supervised k-means as the GCD baseline, this dynamic conditioning yields the reported accuracy gains, with the largest gains in novel-ID accuracy. The paper further claims that the mechanism is parameter-efficient, needing only about 20% trainable parameters versus a static prefix generator's 72%.

Load-bearing premise

The evaluation assumes that the identities to be 'discovered' in YouTube Faces and CASIA are genuinely new to the model, even though the ViT backbones were pretrained on MS1MV3, a large celebrity face collection that may contain the same people; identity overlap between pretraining and evaluation data is never analyzed.

Editorial extensions

If this is right

  • On all six GFD benchmarks, FaceGCD reports the highest All, Known, and Novel accuracy among the compared GCD methods and ArcFace, e.g., 82.7% All / 93.2% Known / 72.1% Novel on YTF 1000.
  • The ablation shows that dynamic, instance-specific prefix generation is essential: a static prefix pool loses 40.3 points overall, and a larger static prefix generator (over 10x parameters) still loses 6.0 points.
  • FaceGCD's features also support generic clustering algorithms: it beats ArcFace for every clustering scheme tested in Table 2, including K-Means, DBSCAN, HAC, L-GCN, Ada-NETS, and SSK.
  • The method transfers to generic GCD benchmarks (CIFAR-100, ImageNet-100, CUB, Stanford Cars, FGVC-Aircraft, Herbarium), achieving state-of-the-art or competitive results especially on fine-grained datasets, without any face-specific design.
  • The approach is parameter-efficient: only 19.8% of total parameters are trainable, versus 72.0% for the static prefix generator variant.

Reading between the lines

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

  • If the dynamic per-instance conditioning is the true cause of the gains, the same hypernetwork-plus-prefix recipe should transfer to other fine-grained, high-cardinality recognition domains (e.g., bird species, retail products, or cell phenotypes), where generic GCD models also struggle.
  • The paper evaluates with the number of clusters set equal to the true number of identities, which is known in these benchmarks; a harder and more realistic test would let the model estimate the cluster count, and the claimed advantage might shrink.
  • Because both ViT backbones are pretrained on face data, the paper cannot cleanly separate the contribution of the dynamic prefix mechanism from the contribution of face-specific pretraining; a disentangled study with ImageNet-only pretraining would settle this.
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 / 7 minor

Summary. The paper introduces a new problem formulation, Generalized Face Discovery (GFD), which unifies known-ID face identification with discovery of previously unseen face identities, and proposes FaceGCD, a method that dynamically generates layer-wise prefix tokens for a frozen ViT backbone using a hypernetwork conditioned on the input image. The method is evaluated on six GFD datasets built from YouTube Faces and CASIA-WebFaces (500, 1000, 2000 identities each), compared against four GCD baselines and against ArcFace with several clustering algorithms, and ablated on prefix size, static versus dynamic prefix generation, and hypernetwork use. The central empirical claim is that FaceGCD outperforms all baselines on all six datasets, with the strongest gain on novel ID accuracy, e.g., 72.1% Novel accuracy on YTF 1000 versus 63.8% for PromptCAL. The paper also reports competitive results on generic GCD benchmarks in the appendix. The authors state that code and data splits are not released, and no statistical significance or multiple-seed results are reported.

Significance. If the empirical results hold, the paper makes two contributions: (i) it proposes a new task, GFD, that is a natural and previously unaddressed combination of face identification and generalized category discovery; and (ii) it demonstrates that dynamic, instance-specific prefix generation via a hypernetwork can substantially beat both generic GCD methods and a strong face recognition baseline on that task. The idea of generating prefix parameters on the fly, rather than using a static prompt pool or a large static generator, is sensible and is supported by a clean ablation (Table 3), where the static prefix pool collapses by 40 points and a static prefix generator underperforms despite much larger capacity. The paper also provides consistent evidence across six datasets and across multiple clustering algorithms, and it performs extra validation on generic GCD benchmarks.

major comments (4)
  1. [Section 4, Implementation Details and Table 1] The central claim that FaceGCD discovers 'new, previously unseen IDs' is not established because the pretraining data and evaluation data may share identities. The manuscript states that both the ViT-B/16 DINO backbone and the ResNet-101 ArcFace model are pretrained on MS1MV3 (Section 4, Implementation Details), while the six GFD datasets are sampled from YTF and CASIA (Appendix B). MS1MV3 is a large celebrity collection and YTF/CASIA are standard celebrity benchmarks, so identity overlap is plausible. If the 'unknown' YTF/CASIA identities appear in MS1MV3, the DINO backbone has already seen images of those identities, and the reported Novel-ID accuracies in Table 1 could reflect retrieval of pretraining memory rather than discovery of genuinely unseen identities. The authors should quantify identity overlap between MS1MV3 and each sampled evaluation split, report results on a subset of identities that are provably disjoint from pretraining, or at minimum include a baseline that removes identities from the pretraining set or tests on a disjoint dataset. Without this, the absolute Novel-ID numbers and the open-world discovery interpretation are not supported, even though relative rankings among ViT-based methods may survive.
  2. [Section 4.1 and Appendix A.4] All experimental conclusions rest on single runs with no error bars or multiple seeds. Table 1 reports point accuracies such as 82.7% for FaceGCD on YTF 1000 and 75.2% for PromptCAL, and Table 3 reports differences as small as 0.2-0.7 points (e.g., prefix size 10 vs 20) as meaningful. Given the standard deviations typical of k-means-based clustering evaluations, differences of this magnitude may not be significant. The authors should report means and standard deviations over at least 3-5 random seeds (or over different identity splits), and ideally include significance tests for the headline comparisons. This is also relevant to the claim in Section 4.2 that performance 'saturates' around prefix size 10, which is currently based on a 0.2 point difference.
  3. [Appendix C.2, Table A5] The generality claim in Appendix C.2 is overstated: on the generic GCD benchmarks, FaceGCD is not state-of-the-art overall. On ImageNet-100, FaceGCD's All accuracy (78.4) is lower than GCD (74.1? actually lower than PromptCAL 83.0 and CMS 84.7) and on CIFAR-100 it is 83.4 vs CMS 82.3, which is within noise, while on Herbarium19 it is notably below several baselines (42.6 vs 56.0 for CMS). The text says 'competitive or superior' and 'state-of-the-art on fine-grained datasets like CUB and Stanford Cars,' but the table shows FaceGCD is best on CUB (64.5) and Stanford Cars (62.5), while on Aircraft and Herbarium19 it is behind CMS. This is not a fatal issue for the main face-domain claim, but the generalization statement in the last paragraph of Appendix C.2 should be revised to accurately report the mixed results, especially on Herbarium19.
  4. [Section 3.2 and Appendix A.2] There is a discrepancy in the stated and used prefix size. Section 3.2 and Appendix A.2 set the prefix size to 20 per layer ('we set the prefix size to 20 per layer'), and Table 3 reports the main FaceGCD model with prefix size 10 (82.9) as the default, with prefix size 20 giving 82.7. The ablation table and the implementation details are inconsistent about which configuration produces the reported Table 1 numbers. The authors should clarify which prefix size is used for each reported result and ensure the main configuration matches the ablation default.
minor comments (7)
  1. [Abstract and Section 1] The phrase 'a key step toward artificial general intelligence (AGI)' in the abstract and introduction is not supported by the experiments and is of limited relevance to the technical content; consider softening or removing it.
  2. [Section 1, Figure 1 caption] There is a typo in the caption: 'Uknown IDs' should be 'Unknown IDs'. Also, Figure 2's left panel mentions 'YTF 1000' but the axis labels are not described; the figure should be readable without zoom.
  3. [Section 3.1] The problem formulation defines the goal as assigning each test image to a known ID or a newly discovered ID, but it does not specify how the number of clusters for unknown IDs is determined. The experiments set the number of clusters to the ground-truth number of classes (Appendix A.3), which is a luxury unavailable in a real open-world setting; this should be stated in the main text, and the sensitivity to the estimated cluster count should be discussed.
  4. [Section 3.3] The loss notation is confusing because tau is used both as a temperature and as a weighting factor in the total loss L_B = (1-tau) sum L_u + lambda sum L_s. In Appendix A.1 the same symbol tau is used as the temperature inside the contrastive losses. The authors should use distinct symbols for the temperature and the unsupervised-loss weight.
  5. [Appendix B, Table A3] The dataset construction section mentions filtering IDs with fewer than 100 images for YTF and 50 for CASIA, but it does not report how many IDs remain after filtering or how the random sampling and train/test splits were seeded. Including these details, or releasing the exact split indices, would improve reproducibility.
  6. [Appendix A.4] The static prefix pool ablation uses a pool of 10 prefixes with 5 selected per instance for 'Static Prefix Pool', while the main model uses 10 or 20 prefixes. The comparison is reasonable, but the text should specify whether the pool size was tuned or chosen to match parameter counts; as written, the 40-point drop could partly reflect a poorly tuned pool size.
  7. [References] Reference [30] is cited as 'SIBGRAPI 2003' in the bibliography while the conference year is 2023; this typo should be corrected. Also, several URLs in the references (e.g., [40]) are given as bare GitHub links without access dates; consider adding them.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirically measured benchmark results, with loss functions and metrics adopted from external prior work rather than derived from the conclusion.

full rationale

FaceGCD is an empirical systems paper, not a derivation. The central claim is benchmark accuracy on held-out test splits (Table 1 and Table 2), and the supporting ablations (Table 3) compare architectural variants with other factors fixed. The loss function is explicitly borrowed from the standard GCD baseline: the paper states 'To highlight the effectiveness of our architectural design, we adopt the same semi-contrastive loss used in the standard GCD baseline [32]'. Adopting a baseline loss is not circular, because the contribution is the architecture, not the loss. Hyperparameters are fixed and selected on validation accuracy ('the best-performing model on validation (All accuracy) is selected for evaluation'), and the evaluation uses the standard Hungarian assignment protocol from the GCD literature. The NNC separability metric in Appendix C.1 is computed on the same embeddings used for clustering, but it is presented as a descriptive diagnostic of embedding quality, not as a prediction or as evidence that is forced by construction; reporting a metric on one's own outputs is not circularity. There are no load-bearing self-citations: the reference list contains no prior work by the present authors, and all cited methods (DINO, GCD, PromptCAL, SimGCD, CMS, ArcFace, etc.) are external. The potential identity overlap between the MS1MV3 pretraining set and the YTF/CASIA evaluation sets is a real data-contamination and validity concern, but it does not make the derivation circular; it concerns whether the benchmark measures what it claims to measure. Accordingly, no circular step can be exhibited with a quote and a specific reduction, and the appropriate score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a standard pretrained backbone, a second frozen ViT for conditioning, a hypernetwork with hand-chosen bottleneck dimensions, and the GCD contrastive loss. The main unstated premise is that the evaluation identities are truly absent from the pretraining data.

free parameters (5)
  • Prefix size = 20 per layer
    Ablated in Table 3; size 20 chosen, with minor sensitivity.
  • Contrastive temperature tau = 1.0
    Taken from prior GCD setups; held fixed.
  • Supervised loss weight lambda = 0.35
    Selected based on validation performance.
  • Hypernetwork bottleneck dimension = 16 (down) and 64 (up)
    Design choice for prefix generator projections; no ablation.
  • SSK nearest-neighbor settings = explored in {10,20,100,200}
    Clustering hyperparameters explored and selected by accuracy.
assumptions (5)
  • domain assumption DINO-pretrained ViT on MS1MV3 provides useful face features
    The method relies on a frozen backbone and a frozen static feature extractor that were pretrained on a large face dataset.
  • domain assumption Landmark-guided patch extraction improves face recognition
    All methods use the MobileNetV3 landmark detector and patch partitioning from Part-based face recognition; no GFD-specific validation.
  • domain assumption Semi-supervised k-means with known number of clusters is a valid evaluation
    The standard GCD evaluation protocol assumes the total number of IDs is known; this is reasonable for benchmarking but is an assumption about the task.
  • domain assumption Evaluation identities are disjoint from pretraining identities
    Unstated in the paper; if false, novel-ID accuracy is inflated.
  • domain assumption The semi-contrastive loss from GCD is sufficient for GFD
    The paper adopts the GCD loss unchanged, so the method's success depends on this loss being adequate for face ID discovery.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FaceGCD: Generalized Face Discovery via Dynamic Prefix Generation." pith.science (2026). https://pith.science/paper/G4PSOQ3Z

@misc{pith2026250722353,
  author       = {Pith},
  title        = {Pith review of: FaceGCD: Generalized Face Discovery via Dynamic Prefix Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G4PSOQ3Z}},
  note         = {Machine review of arXiv:2507.22353}
}
read the original abstract

Recognizing and differentiating among both familiar and unfamiliar faces is a critical capability for face recognition systems and a key step toward artificial general intelligence (AGI). Motivated by this ability, this paper introduces generalized face discovery (GFD), a novel open-world face recognition task that unifies traditional face identification with generalized category discovery (GCD). GFD requires recognizing both labeled and unlabeled known identities (IDs) while simultaneously discovering new, previously unseen IDs. Unlike typical GCD settings, GFD poses unique challenges due to the high cardinality and fine-grained nature of face IDs, rendering existing GCD approaches ineffective. To tackle this problem, we propose FaceGCD, a method that dynamically constructs instance-specific feature extractors using lightweight, layer-wise prefixes. These prefixes are generated on the fly by a HyperNetwork, which adaptively outputs a set of prefix generators conditioned on each input image. This dynamic design enables FaceGCD to capture subtle identity-specific cues without relying on high-capacity static models. Extensive experiments demonstrate that FaceGCD significantly outperforms existing GCD methods and a strong face recognition baseline, ArcFace, achieving state-of-the-art results on the GFD task and advancing toward open-world face recognition.

Figures

Figures reproduced from arXiv: 2507.22353 by the authors.

Figure 1
Figure 1. Generalized Face Discovery. Given a training set of labeled and unlabeled face images, GFD aims to classify known IDs and simultaneously discover unknown IDs at inference time. 1. More specifically, given a dataset con￾sisting of labeled face images of known IDs and unlabeled face images corresponding to either known or unknown IDs, our goal is to train a model that can classify known face images into their correspo… view at source ↗
Figure 2
Figure 2. (Left) Overall performance. Separability (defined in Appendix C.1) and accuracy on YTF 1000 for GCD [32], ArcFace [5], and FaceGCD (ours). (Right) Feature space visualization. t-SNE features of known (0 - 4) and unknown (500 - 504) IDs. as summarized in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. FaceGCD architecture overview. supervised learning setting like GFD. Therefore, it is crucial to devise a more adaptive and ef￾ficient mechanism capable of extracting ID-sensitive features without relying on brute-force model scaling. 3.2 FaceGCD with Dynamic Prefix Generation In this paper, we propose FaceGCD, which effectively addresses the aforementioned chal￾lenges of GFD by employing our novel architecture that… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 37 canonical work pages

  1. [1]

    Emerging properties in self-supervised vision transform- ers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bo- janowski, and Armand Joulin. Emerging properties in self-supervised vision transform- ers. In2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 9630–9640, 2021

    work page 2021

  2. [2]

    Vinod Kumar Chauhan, Jiandong Zhou, Ping Lu, Soheila Molaei, and David A. Clifton. A brief review of hypernetworks in deep learning.Artif. Intell. Rev., 57, 2024

    work page 2024

  3. [3]

    Contrastive mean-shift learning for gener- alized category discovery

    Sua Choi, Dahyun Kang, and Minsu Cho. Contrastive mean-shift learning for gener- alized category discovery. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 23094–23104, 2024

    work page 2024

  4. [4]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, pages 248–255, 2009

    work page 2009

  5. [5]

    Arcface: Additive angular margin loss for deep face recognition.CoRR, abs/1801.07698, 2018

    Jiankang Deng, Jia Guo, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition.CoRR, abs/1801.07698, 2018

    arXiv 2018

  6. [6]

    Lightweight face recognition challenge

    Jiankang Deng, Jia Guo, Debing Zhang, Yafeng Deng, Xiangju Lu, and Song Shi. Lightweight face recognition challenge. In2019 IEEE/CVF International Conference on Computer Vision Workshops, ICCV Workshops 2019, Seoul, Korea (South), October 27-28, 2019, pages 2638–2646, 2019

    work page 2019

  7. [7]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Trans- formers for image recognition at scale. In9th International Conference on Learning Representations, ICLR 202...

    work page 2021

  8. [8]

    A density-based al- gorithm for discovering clusters in large spatial databases with noise

    Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. A density-based al- gorithm for discovering clusters in large spatial databases with noise. InProceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD-96), Portland, Oregon, USA, pages 226–231, 1996

    work page 1996

Show all 44 references
  1. [9]

    Rudd, and Terrance E

    Manuel Günther, Steve Cruz, Ethan M. Rudd, and Terrance E. Boult. Toward open-set face recognition. In2017 IEEE Conference on Computer Vision and Pattern Recogni- tion Workshops, CVPR Workshops 2017, Honolulu, HI, USA, July 21-26, 2017, pages 573–582, 2017

    2017

  2. [10]

    Dai, and Quoc V

    David Ha, Andrew M. Dai, and Quoc V . Le. Hypernetworks. In5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings, 2017

    2017

  3. [11]

    Learning to discover novel visual categories via deep transfer clustering

    Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 8400–8408, 2019. 12YS. O...

    2019

  4. [12]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV , USA, June 27-30, 2016, pages 770–778, 2016

    2016

  5. [13]

    Yun He, Huaixiu Steven Zheng, Yi Tay, Jai Prakash Gupta, Yu Du, Vamsi Aribandi, Zhe Zhao, YaGuang Li, Zhao Chen, Donald Metzler, Heng-Tze Cheng, and Ed H. Chi. Hyperprompt: Prompt-based task-conditioning of transformers. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba ...

    2022

  6. [14]

    Le, Mark Sandler, Bo Chen, Weijun Wang, Liang-Chieh Chen, Mingxing Tan, Grace Chu, Vijay Vasudevan, and Yukun Zhu

    Andrew Howard, Ruoming Pang, Hartwig Adam, Quoc V . Le, Mark Sandler, Bo Chen, Weijun Wang, Liang-Chieh Chen, Mingxing Tan, Grace Chu, Vijay Vasudevan, and Yukun Zhu. Searching for mobilenetv3. In2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Kore...

    2019

  7. [15]

    Belongie, Bharath Hariharan, and Ser-Nam Lim

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Shai Avidan, Gabriel J. Bros- tow, Moustapha Cissé, Giovanni Maria Farinella, and Tal Hassner, editors,Computer Vision - ECCV 2022 - 17th Eur...

    2022

  8. [16]

    Innate face-selectivity in the brain of young domestic chicks.Proceedings of the National Academy of Sciences, 121(40):e2410404121, 2024

    Dmitry Kobylkov, Orsola Rosa-Salva, Mirko Zanon, and Giorgio Vallortigara. Innate face-selectivity in the brain of young domestic chicks.Proceedings of the National Academy of Sciences, 121(40):e2410404121, 2024

    2024

  9. [17]

    3d object representations for fine-grained categorization

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In2013 IEEE International Conference on Computer Vision Workshops, ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 554–561, 2013

    2013

  10. [18]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    2009

  11. [19]

    The hungarian method for the assignment problem.Naval research logistics quarterly, 2(1-2):83–97, 1955

    Harold W Kuhn. The hungarian method for the assignment problem.Naval research logistics quarterly, 2(1-2):83–97, 1955

    1955

  12. [20]

    The power of scale for parameter- efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter- efficient prompt tuning. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih, editors,Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processin...

    2021

  13. [21]

    Prefix-tuning: Optimizing continuous prompts for generation

    Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Chengqing Zong, Fei Xia, Wenjie Li, and Roberto Navigli, editors, Proceedings of the 59th Annual Meeting of the Association for Computational Lin- guistics and the 11th International...

    2021

  14. [22]

    Sphereface: Deep hypersphere embedding for face recognition

    Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Sphereface: Deep hypersphere embedding for face recognition. In2017 IEEE Con- ference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 6738–6746, 2017

    2017

  15. [23]

    Stuart P. Lloyd. Least squares quantization in PCM.IEEE Trans. Inf. Theory, 28: 129–136, 1982

    1982

  16. [24]

    Blaschko, and Andrea Vedaldi

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft.CoRR, abs/1306.5151, 2013

    2013 arXiv

  17. [25]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blon- del, P. Prettenhofer, R. Weiss, V . Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Lear...

    2011

  18. [26]

    Local connectivity-based density estimation for face clustering

    Junho Shin, Hyo-Jun Lee, Hyunseop Kim, Jong-Hyeon Baek, Daehyun Kim, and Yeong Jun Koh. Local connectivity-based density estimation for face clustering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Van- couver, BC, Canada, June 17-24, 2023, pag...

    2023

  19. [27]

    SLINK: an optimally efficient algorithm for the single-link cluster method.Comput

    Robin Sibson. SLINK: an optimally efficient algorithm for the single-link cluster method.Comput. J., 16:30–34, 1973

    1973

  20. [28]

    Part-based face recognition with vision transformers

    Zhonglin Sun and Georgios Tzimiropoulos. Part-based face recognition with vision transformers. In33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022, page 611, 2022

    2022

  21. [29]

    Belongie

    Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge J. Belongie. The herbarium challenge 2019 dataset.CoRR, abs/1906.05372, 2019

    2019 arXiv

  22. [30]

    Open-set face recognition with neural ensemble, maximal entropy loss and feature augmentation

    Rafael Henrique Vareto, Manuel Günther, and William Robson Schwartz. Open-set face recognition with neural ensemble, maximal entropy loss and feature augmentation. In36th SIBGRAPI Conference on Graphics, Patterns and Images, SIBGRAPI 2003, Rio Grande, RS, Brazil, November 6-9,...

    2003

  23. [31]

    Boult, William Robson Schwartz, and Manuel Günther

    Rafael Henrique Vareto, Yu Linghu, Terrance E. Boult, William Robson Schwartz, and Manuel Günther. Open-set face recognition with maximal entropy and objectosphere loss.Image Vis. Comput., 141:104862, 2024

    2024

  24. [32]

    Generalized category discovery

    Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 7482–7491, 2022

    2022

  25. [33]

    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. 2011

    2011

  26. [34]

    Cosface: Large margin cosine loss for deep face recognition

    Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018...

    2018

  27. [35]

    Ada-nets: Face clustering via adaptive neighbour discovery in the structure space

    Yaohua Wang, Yaobin Zhang, Fangyi Zhang, Senzhang Wang, Ming Lin, YuQi Zhang, and Xiuyu Sun. Ada-nets: Face clustering via adaptive neighbour discovery in the structure space. InThe Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-...

    2022

  28. [36]

    Linkage based face clus- tering via graph convolution network

    Zhongdao Wang, Liang Zheng, Yali Li, and Shengjin Wang. Linkage based face clus- tering via graph convolution network. InIEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 1117–1125. Computer Vision Foundation / ...

    2019

  29. [37]

    Dy, and Tomas Pfister

    Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer G. Dy, and Tomas Pfister. Dual- prompt: Complementary prompting for rehearsal-free continual learning. In Shai Avi- dan, Gabriel J. Brostow, Moustapha ...

    2022

  30. [38]

    Dy, and Tomas Pfister

    Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guo- long Su, Vincent Perot, Jennifer G. Dy, and Tomas Pfister. Learning to prompt for continual learning. InIEEE/CVF Conference on Computer Vision and Pattern Recog- nition, CVPR 2022, New Orleans, LA, ...

    2022

  31. [39]

    Parametric classification for generalized category discovery: A baseline study

    Xin Wen, Bingchen Zhao, and Xiaojuan Qi. Parametric classification for generalized category discovery: A baseline study. InIEEE/CVF International Conference on Com- puter Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 16544–16554, 2023

    2023

  32. [40]

    Pytorch image models.https://github.com/rwightman/ pytorch-image-models

    Ross Wightman. Pytorch image models.https://github.com/rwightman/ pytorch-image-models

  33. [41]

    Face recognition in unconstrained videos with matched background similarity

    Lior Wolf, Tal Hassner, and Itay Maoz. Face recognition in unconstrained videos with matched background similarity. InThe 24th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2011, Colorado Springs, CO, USA, 20-25 June 2011, pages 529–534, 2011

    2011

  34. [42]

    Dong Yi, Zhen Lei, Shengcai Liao, and Stan Z. Li. Learning face representation from scratch.CoRR, abs/1411.7923, 2014

    2014 arXiv

  35. [43]

    Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fahad Shahbaz Khan

    Sheng Zhang, Salman H. Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fahad Shahbaz Khan. Promptcal: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery. InIEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, CVPR ...

    2023

  36. [44]

    Adacos: Adap- tively scaling cosine logits for effectively learning deep face representations

    Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hongsheng Li. Adacos: Adap- tively scaling cosine logits for effectively learning deep face representations. InIEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages ...

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.