Pith. sign in

REVIEW 5 major objections 6 minor 49 references

Object-level Self-Distillation for Vision Pretraining

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that distilling representations per object, through a mask-guided [OBJ] token and object-aware cropping, improves frozen ViT features at both image and patch level.

desk verdict ODIS is a real step forward for object-level SSL, but the headline DINOv2 comparison is unfair and the paper would be stronger with that claim fixed. read the letter →

arxiv 2506.05409 v1 pith:VV3GNIWC submitted 2025-06-04 cs.CV cs.LG

Çağlar H{i}zl{i} , Çağatay Y{i}ld{i}z , Pekka Marttinen This is my paper
classification cs.CVcs.LG
keywords self-supervisedlearningvisiontransformersobject-leveldistillationsegmentationmasksmaskedattentionImageNetk-NNscene-centricpretrainingdensenearestneighborretrieval
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

State-of-the-art visual pretraining typically distills one vector from a whole image, an implicit single-object assumption that breaks on multi-object and scene-centric data. This paper introduces ODIS, which instead distills an object token that attends only to patches inside a segmentation mask, and it uses object-aware cropping to make sure teacher and student see the same object. The result is better frozen representations: 82.6% k-NN on ImageNet-1k with ViT-Large when ground-truth masks are used at inference, along with consistent gains in linear probing and patch-level dense retrieval. The paper's broader claim is that object-level distillation is a scalable alternative to image-level distillation for pretraining on uncurated, multi-object datasets.

What carries the argument

The central mechanism is the [OBJ] token combined with masked multi-head attention. A ViT receives patch tokens plus a learnable object token; at every layer, that token's attention is restricted to patches where a binary object mask $\tilde{y}$ is 1, so its output is a nonlinear, mask-conditioned pooling of object patches. Object-aware cropping randomly resamples global views until the target object appears in both teacher and student crops, and when an image contains several objects a single target is sampled per forward pass (uniformly or by area). The object-level loss is the cross-entropy between teacher and student softmax distributions over the [OBJ] token, added to iBOT's masked patch prediction loss.

What would settle it

Evaluate a ViT-Large ODIS model pretrained with only predicted masks and tested with the same predicted masks on a multi-object benchmark; if its k-NN accuracy no longer beats iBOT run under identical mask conditions, the reported gains come from mask quality rather than from object-level distillation itself.

Watch

Extended reading notes

Core claim

ODIS takes the teacher-student self-distillation recipe and changes what is being distilled. Instead of a [CLS] token that pools over the whole image, it inserts a [OBJ] token whose attention is masked so it can only aggregate patches inside a binary object segmentation map. The teacher and student receive global crops that are resampled until they both contain the target object, and the training loss is the cross-entropy between their [OBJ] distributions plus iBOT's masked patch prediction loss. The paper reports that this object-level objective improves both image-level benchmarks (k-NN and linear probing on ImageNet-1k) and patch-level benchmarks (dense nearest-neighbor retrieval on PASCAL VOC and ADE20k), and that the improvements hold even when masks are not used at inference, though they are largest when masks are available.

Load-bearing premise

The method assumes a binary object mask can reliably mark the target object during both pretraining and inference; with missing, wrong, or only-main-object masks the object-aware crops and masked attention lose their advantage, and the headline numbers use clean ground-truth ImageNet masks.

Editorial extensions

If this is right

  • ODIS can answer object queries, not just image queries: the same frozen backbone can produce a representation for any object whose mask is supplied at inference.
  • Pretraining on scene-centric COCO with ODIS beats iBOT by 4.2 points in ImageNet k-NN, suggesting object-level objectives will scale better to uncurated multi-object data.
  • Masks from off-the-shelf segmenters are already good enough to improve pretraining over iBOT, so the method does not require clean ground-truth masks to be useful.
  • Because the object-level loss replaces the image-level loss in the final objective, ODIS shows that a purely object-centric distillation signal is sufficient; no image-level term is needed.

Reading between the lines

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

  • The mask-guided pooling recipe could be applied to downstream object-centric tasks such as referring segmentation or visual question answering, where a query object representation is needed; the paper does not test these.
  • The finding that local crops should be drawn from context-rich regions rather than object windows suggests object-in-context is the productive pretraining regime, a hypothesis not fully isolated by the ablations.
  • A natural large-scale test is ODIS with automatically generated masks on a multi-object corpus at DINOv2-scale data and compute; the paper only provides small-model evidence for external masks.
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

5 major / 6 minor

Summary. ODIS proposes to replace the image-level [CLS] self-distillation objective of DINO/iBOT with an object-level [OBJ] token, using object-aware cropping and masked attention so that teacher and student embed the same target object, while keeping iBOT's patch-level masked prediction loss. The paper reports ImageNet-1k k-NN and linear probing results for ViT-S/B/L pretrained on ImageNet-1k and ViT-S pretrained on COCO, dense nearest-neighbor retrieval results on PASCAL VOC and ADE20k, ablations of loss components, cropping configurations, object sampling, and mask sources, and a runtime comparison. The central claim is that object-level distillation improves both image- and patch-level representations, and that when ground-truth masks are used at inference, ODIS with ViT-L reaches 82.6% k-NN on ImageNet-1k, which is +0.6 over a reported DINOv2 baseline.

Significance. If the training-time benefit is robust, ODIS is a conceptually simple and broadly applicable modification: it converts segmentation masks, which are often available or cheap to obtain, into an object-level training signal, and it also works with bounding boxes. The no-mask comparisons in Table 1 (e.g., ViT-L 79.6 vs 78.0 for iBOT; ViT-B 78.3 vs 77.1 for iBOT) and the dense-retrieval gains in Table 3 are promising and support a real training-time improvement. The ablations in Tables 4 and 5, especially the comparison of mask sources and the loss-component study, are useful and give the paper empirical substance. The significance is currently limited by two facts: the headline comparison to DINOv2 is not controlled for the inference-time mask protocol, and the paper does not provide any variance or multiple-seed evidence for claims that rest on small margins. No machine-checked proofs or released code are provided, but the experimental protocol is described in enough detail to replicate most parts once the mask source is clarified.

major comments (5)
  1. [Section 5.1, Table 1] The DINOv2 comparison is not controlled: the ODIS+Masks row for ViT-L reports 82.6% k-NN using ground-truth ImageNet masks at inference, while the DINOv2-Sc. ViT-L row is evaluated without masks. The authors explicitly acknowledge in Section 5.1 that 'applying segmentation masks at inference would likely improve DINOv2's performance,' yet they do not run that experiment. Because the same masked-attention modification used to produce the DINO/iBOT+Masks rows can be applied to the public DINOv2 checkpoint, this control is feasible. Without it, the abstract's '+0.6% over DINOv2' claim is unsupported and could be a protocol artifact.
  2. [Table 1 and Abstract] The arrows in Table 1 and the abstract's '+4.6% improvement over iBOT' compare ODIS+Masks against iBOT without masks (78.0 for ViT-L), while the body text reports gains of +2.3 to +2.7 against iBOT+Masks in the green-ticked rows. This conflates the training-time benefit of ODIS with the test-time use of oracle masks. Please report every delta against the matching inference protocol and state explicitly in the table caption and abstract which baseline is used for each arrow.
  3. [Section 5.1, paragraph 'Next, we compare against DINOv2'] The paper uses a DINOv2 ViT-L number that the authors state they failed to reproduce: 'our attempts to fully replicate DINOv2 ViT-L results were unsuccessful.' The reported 82.0/84.5 for DINOv2-Sc. ViT-L is therefore an external literature number, and a +0.6 margin over an unreproduced baseline is not a reliable quantitative claim. Either verify the baseline with the authors' own evaluation pipeline or clearly label the comparison as literature-reported and downgrade the claim accordingly.
  4. [Section 4, Implementation details, 'Object segmentation maps'] The paper says 'For IN1k, a single object segmentation map is provided for each image,' but the standard ImageNet-1k release provides bounding boxes for the localization task, not segmentation maps. The manuscript must specify the exact annotation source and format for the ImageNet masks, including whether they are boxes converted to binary maps. This matters because masked attention over a bounding-box region still pools over background inside the box, which weakens the claimed removal of background 'free-riders' and affects the interpretation of Table 4.
  5. [Tables 1-5] All reported numbers appear to come from a single pretraining run per configuration. The differences used to support the central claims are sometimes small (e.g., +0.6 over DINOv2, or +1.0 to +2.6 mIoU in Table 3), and no error bars, seeds, or statistical tests are provided. Please provide at least three pretraining seeds for the smallest model per setting, or otherwise justify that the reported differences are above run-to-run variance.
minor comments (6)
  1. [Section 4, 'Masked attention'] The paper does not specify whether patch tokens are allowed to attend to the [OBJ] token in the masked attention matrix; clarify which rows and columns of the attention matrix are masked, since this changes the exact implementation and the information flow from background patches.
  2. [Section 4, inference without masks] When no segmentation mask is available at inference, the manuscript should state explicitly that the [OBJ] token uses full attention (equivalently an all-ones mask); currently the reader has to infer this behavior from the 'whole image' assumption for missing training masks.
  3. [Table 4 and related text] The rows labeled 'YOLO' and 'MA VL' refer to object detectors that output bounding boxes, not segmentation maps; rename these rows or consistently say 'bounding boxes' so that the comparison in Table 4 is not misleading.
  4. [Table 1 caption] The DINOv2 rows do not state whether the reported k-NN and LP numbers were re-evaluated by the authors or taken from the DINOv2 paper; add this information and specify the evaluation protocol used for those rows.
  5. [Section 5.1, Figure 4] For the iBOT 'Masks' attention maps, it is unclear whether the [CLS] token is used with masked attention and how the mask is applied to a checkpoint that was not trained with an [OBJ] token; please describe the exact inference procedure for the baseline +Masks rows.
  6. [Abstract] The phrase '82.6% k-NN accuracy on ImageNet1k with ViT-Large' should specify ViT-Large/16 and the pretraining dataset (ImageNet-1k) to avoid ambiguity with the larger ImageNet-22k or other patch sizes.

Circularity Check

1 steps flagged · score 2.0 of 10

No derivation-level circularity; the core objective is independently benchmarked, but the headline DINOv2 comparison uses an unequal inference protocol that the paper itself concedes could favor ODIS.

  1. other [Section 5.1 (Table 1 and DINOv2 comparison paragraph)]
    "Our experiments show that ODIS surpasses the DINOv2 ViT-L model trained on IN22k by +0.6 percentage points in k-NN classification accuracy. While it is true that applying segmentation masks at inference would likely improve DINOv2’s performance, ODIS is expected to similarly benefit from scaling up the model and data, as well as the same set of algorithmic advances."

    The claimed +0.6 is not a like-for-like comparison of inference procedures: the ODIS+Masks ViT-L row (82.6) is obtained by feeding the [OBJ] token ground-truth ImageNet object masks at inference, whereas the DINOv2-Sc. ViT-L row (82.0) is evaluated with the standard no-mask protocol. ODIS is therefore given object-location information that DINOv2 is denied, and the paper explicitly concedes that masks would likely improve DINOv2. The headline gap may thus be a protocol artifact rather than evidence for the object-level pretraining objective. This is a circular-comparison caveat, not a derivation-level circularity: the no-mask ODIS-vs-iBOT rows and the DINO/iBOT+Masks control rows provide independent evidence for the core method.

full rationale

The central derivation is self-contained: ODIS's objective LODIS = L[OBJ] + L[PATCH] (Eq. 14) combines object-level distillation with iBOT-style patch loss, and its reported k-NN, LP, and mIoU figures are measured against external benchmarks (ImageNet labels, PASCAL VOC, ADE20k). No parameter is fitted to those benchmarks, and no load-bearing claim rests on the authors' own prior theorems; DINO, iBOT, and DINOv2 are third-party baselines. The 'DINO+Masks' and 'iBOT+Masks' rows control for the inference-time masked-attention protocol, and the no-mask ODIS rows show a training-time gain over iBOT (e.g., 79.6 vs 78.0 for ViT-L), so the core method has independent content. The flagged limitation is the DINOv2 comparison: ODIS's 82.6 k-NN uses ground-truth masks at inference while DINOv2's 82.0 does not, and the paper admits that applying masks would likely improve DINOv2. That makes the headline +0.6 claim vulnerable to a protocol artifact, but it is a fairness/evaluation gap rather than a derivation that reduces to its inputs. Overall circularity is minimal.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The method introduces no fitted numerical parameters beyond conventional hyperparameters; it introduces no new physical or conceptual entities. Its central assumption is the availability and correctness of object masks, which is explicit in the paper. The [OBJ] token is a model component, not an invented external entity.

free parameters (3)
  • Teacher/student softmax temperatures (t_s, t_t) = not reported; inherited from DINO/iBOT
    Temperature hyperparameters in the distillation cross-entropy loss; chosen by hand in prior work, not optimized here.
  • Maximum object-aware crop attempts = 20
    Section 4, '1 Object-aware cropping': 'randomly cropping up to 20 times and keeping the global views that contain the target object.' This is a hand-picked constant.
  • Patch masking ratio p = p ~ U[0.1, 0.5]
    Section C: random block masking with p sampled uniformly between 0.1 and 0.5, inherited from iBOT. It affects the patch-level loss strength.
assumptions (3)
  • domain assumption Ground-truth or generated object segmentation maps are available for pretraining images and inference images.
    Section 4 ('1 Object-aware cropping') and Section 5.1 ('Scenario-1: segmentation masks available during inference'). The entire method and the headline 82.6% number depend on this.
  • domain assumption A single target object per forward pass is sufficient; sampling one object at a time yields improvements.
    Section 4 ('we sample a single target object per forward pass') and ablations in Table 5 compare random vs area-proportional sampling, but always a single object.
  • standard math Standard self-distillation with EMA teacher (DINO/iBOT) provides a stable training signal for object tokens.
    Section 3 and Section 4; the loss and EMA update are taken directly from prior work and are not re-derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Object-level Self-Distillation for Vision Pretraining." pith.science (2026). https://pith.science/paper/VV3GNIWC

@misc{pith2026250605409,
  author       = {Pith},
  title        = {Pith review of: Object-level Self-Distillation for Vision Pretraining},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VV3GNIWC}},
  note         = {Machine review of arXiv:2506.05409}
}
abstract

State-of-the-art vision pretraining methods rely on image-level self-distillation from object-centric datasets such as ImageNet, implicitly assuming each image contains a single object. This assumption does not always hold: many ImageNet images already contain multiple objects. Further, it limits scalability to scene-centric datasets that better mirror real-world complexity. We address these challenges by introducing Object-level Self-DIStillation (ODIS), a pretraining approach that shifts the self-distillation granularity from whole images to individual objects. Using object-aware cropping and masked attention, ODIS isolates object-specific regions, guiding the transformer toward semantically meaningful content and transforming a noisy, scene-level task into simpler object-level sub-tasks. We show that this approach improves visual representations both at the image and patch levels. Using masks at inference time, our method achieves an impressive $82.6\%$ $k$-NN accuracy on ImageNet1k with ViT-Large.

Figures

Figures reproduced from arXiv: 2506.05409 by the authors.

Figure 1
Figure 1. (Left) Multi-object example from Im￾ageNet. Taken from [Yun et al., 2021]. (Right) Teacher and student see crops of distinct objects. At the image-level, self-distillation is im￾plemented via a single-label classification objective on a global [CLS] embedding. While effective for object-centric datasets such as ImageNet [Deng et al., 2009], it implicitly assumes that each image contains a single object and the singl… view at source ↗
Figure 2
Figure 2. Image-level vs. Object-level distillation. (a) Standard random cropping have no inherent mechanism to ensure that the student and teacher receive the same object as input. Hence, the distilled [CLS] tokens may summarize semantically different entities. (b) Our approach resolves this issue by 1 Object-aware Cropping that uses object masks. Further, 2 Masked Attention guides the [OBJ] token to pool information only fr… view at source ↗
Figure 3
Figure 3. Masked Attention with Object Segmentation Masks. Next, we detail Object-level Self-Distillation (ODIS), our pro￾posed pretraining method that redefines self-distillation at the object level rather than the conventional image level. ODIS is built around two key components: 1 object-aware cropping, which ensures that both student and teacher networks receive dis￾tinct views of the same object, and 2 masked attention, … view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: An example input, ODIS and iBOT attention maps using [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: An extended version of Fig. 4, where all attention heads are visualized. [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Additional examples showing iBOT’s failure despite masked attention. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Examples showing how iBOT fails in retrieving a nearest neighbor with the correct class [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 21 canonical work pages

  1. [1]

    Deep vit features as dense visual descriptors

    Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2 0 (3): 0 4, 2021

    arXiv 2021

  2. [2]

    Self-supervised learning from images with a joint-embedding predictive architecture

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619--15629, 2023

    2023

  3. [3]

    Towards in-context scene understanding

    Ivana Balazevic, David Steiner, Nikhil Parthasarathy, Relja Arandjelovi \'c , and Olivier Henaff. Towards in-context scene understanding. Advances in Neural Information Processing Systems, 36, 2024

    work page 2024

  4. [4]

    Beit: Bert pre-training of image transformers

    Hangbo Bao, Li Dong, Fuliang Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021

    arXiv 2021

  5. [5]

    Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020

    Lucas Beyer, Olivier J H \'e naff, Alexander Kolesnikov, Xiaohua Zhai, and A \"a ron van den Oord. Are we done with imagenet? arXiv preprint arXiv:2006.07159, 2020

    arXiv 2006

  6. [6]

    Monet: Unsupervised scene decomposition and representation

    Christopher P Burgess, Loic Matthey, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. Monet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019

    arXiv 1901

  7. [7]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213--229. Springer, 2020

    2020

  8. [8]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv \'e J \'e gou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650--9660, 2021

    2021

Show all 49 references
  1. [9]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597--1607. PmLR, 2020

    2020

  2. [10]

    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. In 2009 IEEE conference on computer vision and pattern recognition, pages 248--255. Ieee, 2009

    2009

  3. [11]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

    2018 arXiv

  4. [12]

    On the transfer of object-centric representation learning

    Aniket Rajiv Didolkar, Andrii Zadaianchuk, Anirudh Goyal, Michael Curtis Mozer, Yoshua Bengio, Georg Martius, and Maximilian Seitzer. On the transfer of object-centric representation learning. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  5. [13]

    Attention over learned object embeddings enables complex visual reasoning

    David Ding, Felix Hill, Adam Santoro, Malcolm Reynolds, and Matt Botvinick. Attention over learned object embeddings enables complex visual reasoning. Advances in neural information processing systems, 34: 0 9112--9124, 2021

    2021

  6. [14]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv...

    2010 arXiv

  7. [15]

    The pascal visual object classes challenge: A retrospective

    Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. International journal of computer vision, 111: 0 98--136, 2015

    2015

  8. [16]

    Bootstrap your own latent-a new approach to self-supervised learning

    Jean-Bastien Grill, Florian Strub, Florent Altch \'e , Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neur...

    2020

  9. [17]

    Unsupervised semantic segmentation by distilling feature correspondences

    Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman. Unsupervised semantic segmentation by distilling feature correspondences. arXiv preprint arXiv:2203.08414, 2022

    2022 arXiv

  10. [18]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll \'a r, and Ross Girshick. Masked autoencoders are scalable vision learners. CVPR, 2022

    2022

  11. [19]

    Efficient visual pretraining with contrastive detection

    Olivier J H \'e naff, Skanda Koppula, Jean-Baptiste Alayrac, Aaron Van den Oord, Oriol Vinyals, and Joao Carreira. Efficient visual pretraining with contrastive detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10086--10096, 2021

    2021

  12. [20]

    Object discovery and representation networks

    Olivier J H \'e naff, Skanda Koppula, Evan Shelhamer, Daniel Zoran, Andrew Jaegle, Andrew Zisserman, Jo \ a o Carreira, and Relja Arandjelovi \'c . Object discovery and representation networks. In European Conference on Computer Vision, pages 123--143. Springer, 2022

    2022

  13. [21]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015--4026, 2023

    2023

  14. [22]

    Cribo: Self-supervised learning via cross-image object-level bootstrapping

    Tim Lebailly, Thomas Stegm \"u ller, Behzad Bozorgtabar, Jean-Philippe Thiran, and Tinne Tuytelaars. Cribo: Self-supervised learning via cross-image object-level bootstrapping. arXiv preprint arXiv:2310.07855, 2023

    2023 arXiv

  15. [23]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll \'a r, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision--ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, procee...

    2014

  16. [24]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38--55. Springer, 2024

    2024

  17. [25]

    Object-centric learning with slot attention

    Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. Advances in neural information processing systems, 33: 0 11525--11538, 2020

    2020

  18. [26]

    Class-agnostic object detection with multi-modal transformer

    Muhammad Maaz, Hanoona Rasheed, Salman Khan, Fahad Shahbaz Khan, Rao Muhammad Anwer, and Ming-Hsuan Yang. Class-agnostic object detection with multi-modal transformer. In European conference on computer vision, pages 512--531. Springer, 2022

    2022

  19. [27]

    Exploring the effectiveness of object-centric representations in visual question answering: Comparative insights with foundation models

    Amir Mohammad Karimi Mamaghan, Samuele Papa, Karl Henrik Johansson, Stefan Bauer, and Andrea Dittadi. Exploring the effectiveness of object-centric representations in visual question answering: Comparative insights with foundation models. arXiv preprint arXiv:2407.15589, 2024

    2024 arXiv

  20. [28]

    Unsupervised learning of dense visual representations

    Pedro O O Pinheiro, Amjad Almahairi, Ryan Benmalek, Florian Golemo, and Aaron C Courville. Unsupervised learning of dense visual representations. Advances in neural information processing systems, 33: 0 4489--4500, 2020

    2020

  21. [29]

    Neural congealing: Aligning images to a joint semantic atlas

    Dolev Ofri-Amar, Michal Geyer, Yoni Kasten, and Tali Dekel. Neural congealing: Aligning images to a joint semantic atlas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19403--19412, 2023

    2023

  22. [30]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth \'e e Darcet, Th \'e o Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  23. [31]

    Improving language understanding by generative pre-training.(2018), 2018

    Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training.(2018), 2018

    2018

  24. [32]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pa...

    2021

  25. [33]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R \"a dle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024

    2024 arXiv

  26. [34]

    Do imagenet classifiers generalize to imagenet? In International conference on machine learning, pages 5389--5400

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International conference on machine learning, pages 5389--5400. PMLR, 2019

    2019

  27. [35]

    You only look once: Unified, real-time object detection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779--788, 2016

    2016

  28. [36]

    Are we done with object-centric learning? arXiv preprint arXiv:2504.07092, 2025

    Alexander Rubinstein, Ameya Prabhu, Matthias Bethge, and Seong Joon Oh. Are we done with object-centric learning? arXiv preprint arXiv:2504.07092, 2025

    2025 arXiv

  29. [37]

    Bridging the gap to real-world object-centric learning

    Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Sch \"o lkopf, Thomas Brox, et al. Bridging the gap to real-world object-centric learning. arXiv preprint arXiv:2209.14860, 2022

    2022 arXiv

  30. [38]

    Evaluating machine accuracy on imagenet

    Vaishaal Shankar, Rebecca Roelofs, Horia Mania, Alex Fang, Benjamin Recht, and Ludwig Schmidt. Evaluating machine accuracy on imagenet. In International Conference on Machine Learning, pages 8634--8644. PMLR, 2020

    2020

  31. [39]

    Croc: Cross-view online clustering for dense visual representation learning

    Thomas Stegm \"u ller, Tim Lebailly, Behzad Bozorgtabar, Tinne Tuytelaars, and Jean-Philippe Thiran. Croc: Cross-view online clustering for dense visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7000--7...

    2023

  32. [40]

    Convnets and imagenet beyond accuracy: Understanding mistakes and uncovering biases

    Pierre Stock and Moustapha Cisse. Convnets and imagenet beyond accuracy: Understanding mistakes and uncovering biases. In Proceedings of the European conference on computer vision (ECCV), pages 498--512, 2018

    2018

  33. [41]

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017

    2017

  34. [42]

    From imagenet to image classification: Contextualizing progress on benchmarks

    Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Andrew Ilyas, and Aleksander Madry. From imagenet to image classification: Contextualizing progress on benchmarks. In International Conference on Machine Learning, pages 9625--9635. PMLR, 2020

    2020

  35. [43]

    Splicing vit features for semantic appearance transfer

    Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748--10757, 2022

    2022

  36. [44]

    Dense contrastive learning for self-supervised visual pre-training

    Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024--3033, 2021

    2021

  37. [45]

    Self-supervised visual representation learning with semantic grouping

    Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, and Xiaojuan Qi. Self-supervised visual representation learning with semantic grouping. Advances in neural information processing systems, 35: 0 16423--16438, 2022

    2022

  38. [46]

    Unsupervised object-level representation learning from scene images

    Jiahao Xie, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Unsupervised object-level representation learning from scene images. Advances in Neural Information Processing Systems, 34: 0 28864--28876, 2021

    2021

  39. [47]

    Re-labeling imagenet: from single to multi-labels, from global to localized labels

    Sangdoo Yun, Seong Joon Oh, Byeongho Heo, Dongyoon Han, Junsuk Choe, and Sanghyuk Chun. Re-labeling imagenet: from single to multi-labels, from global to localized labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2340--2350, 2021

    2021

  40. [48]

    Scene parsing through ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633--641, 2017

    2017

  41. [49]

    ibot: Image bert pre-training with online tokenizer

    Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021

    2021 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.