Self-supervised structured object representation learning

Adel Hafiane; Antoine Letienne; Mohamed Abbas Hedjazi; Oussama Hadjerci

REVIEW 5 major objections 6 minor 24 references

Self-supervised structured object representation learning

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Self-supervised ProtoScale builds object-centric representations from semantic grouping, instance separation, and hierarchy, and reports detection gains over DINOv2 and SlotCon under limited labels and epochs.

desk verdict ProtoScale: a plausible SlotCon extension whose own Table 1 contradicts the abstract's 'consistent outperformance' at the lowest data setting. read the letter →

arxiv 2508.19864 v1 pith:6GBK5GSI submitted 2025-08-27 cs.CV

Oussama Hadjerci , Antoine Letienne , Mohamed Abbas Hedjazi , Adel Hafiane This is my paper

classification cs.CV

keywords self-supervisedlearningobject-centricrepresentationsemanticgroupingmulti-scaleobjectdetectiondensepredictionstudent–teacherProtoScale

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 sets out to show that self-supervised visual representations can be structured rather than global: instead of encoding an image as a single vector, the model should group pixels into semantic regions, separate those regions into instance-level units, and relate those units hierarchically. To do this it introduces ProtoScale, a multi-scale grouping module attached to a lightweight ResNet/Transformer pyramid encoder, trained with a student–teacher objective whose losses align semantic, instance, and hierarchical attention maps between augmented views. The paper argues that preserving full scene context across views—rather than relying on random cropping as DINO does—is what makes the learned features useful for dense prediction. On a combined COCO/UA-DETRAC detection benchmark, it reports that ProtoScale outperforms DINOv2, DINOv1, iBOT, TEC, and SlotCon when fine-tuned with the RT-DETR detector, including with only 50% of labeled data and 10 fine-tuning epochs. If correct, this would mean object-centric SSL can be made cheap and data-efficient enough for practical detection in low-resource settings.

What carries the argument

ProtoScale, a multi-level grouping module. It operates on three spatial scales of a ResNet/Transformer pyramid encoder: first, dot-product attention between spatial features and 256 learnable semantic prototypes produces semantic attention maps, regularized by a centered Gaussian prior and supported by auxiliary register prototypes that absorb noise; second, semantic prototypes are softly assigned to 32 batch-shared instance prototypes, converting semantic maps into instance-level attention; third, an MLP predicts an affinity matrix among instance prototypes, thresholded at 0.5, to merge related instances into hierarchical object-part groups. Student and teacher attention maps are aligned by

What would settle it

Fine-tune the same RT-DETR detector from a ProtoScale backbone and from DINOv2's backbone for 10 epochs on 50% of the COCO/UA-DETRAC subset, but train each SSL method with its own original data-augmentation and pretraining setups while controlling total pre-training compute. If the ProtoScale AP gap over DINOv2 shrinks to near zero, the paper's claim that structured grouping causes the gain is falsified; if the gap persists, it supports the claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that a self-supervised objective built from semantic prototypes, batch-shared instance prototypes, and pairwise instance affinities can produce object-centric features that transfer to dense prediction better than global-embedding SSL, with a 12-million-parameter hybrid encoder pretrained for only 35 epochs on COCO. It reports AP 57.1 with 10 fine-tuning epochs on 50% of the labeled COCO/UA-DETRAC data, versus 50.7 for DINOv2, and similar or larger gains at 30% and 70% data, while DINOv2 leads only on large objects. The attribution is to multi-scale semantic grouping with a Gaussian prior and auxiliary register prototypes, batch-level instance grouping, and full-

Load-bearing premise

The load-bearing assumption is that all SSL baselines are compared fairly once each backbone is wrapped in the same shared projection module and trained in the same student–teacher protocol; if some backbones are better or worse suited to that wrapper than ProtoScale is, the reported advantages could come from the wrapper rather than from the learned representations.

Editorial extensions

If this is right

Fine-tuning a ProtoScale-pretrained backbone for 10 epochs with 50% labeled COCO/UA-DETRAC data yields AP 57.1, 6.4 points above DINOv2 under the same detection fine-tuning, while using a 12M-parameter encoder instead of DINOv2's 21M ViT-S.
Preserving full scene context across augmented views can be sufficient for strong dense-prediction transfer, questioning the need for random-crop global views in SSL.
Semantic, instance, and hierarchical grouping can be learned jointly in one SSL pass from unlabeled COCO data, without pseudo-labels or iterative object discovery.
The learned representations transfer particularly well to small and medium objects, where global-embedding SSL methods lose the most accuracy.

Reading between the lines

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

Editorial inference: the largest reported gap appears on small objects (AP_s 31.2 vs 17.9 at 50% data), suggesting the multi-scale grouping is doing real work at fine resolutions; a targeted experiment varying the number of encoder scales would make that mechanism visible.
Editorial inference: because the paper only varies fine-tuning data and epochs, not pre-training data size, a fair test of whether ProtoScale's gains survive at ImageNet/LVD scale remains open; the current comparison mixes pretraining datasets of very different sizes.
Editorial inference: the same attention maps could be probed for part-whole consistency; if instances merge via the affinity threshold, the model should produce stable object-part splits across views, a property not measured by detection AP alone.

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.

Desk Editor's Note

read the letter

Punchline: ProtoScale is a sensible extension of SlotCon that adds batch-level instance prototypes and a hierarchical affinity module, plus a no-crop augmentation strategy. If the numbers are right, it is a useful compute-efficient SSL recipe for downstream detection. But the paper overclaims in the abstract, and the evaluation has enough under-specification that the central claim needs serious revision before it can be trusted.

What is genuinely new: the specific combination of multi-scale semantic prototypes, instance-level prototypes shared across a batch, and hierarchical affinities is not in SlotCon or DINO. The choice to avoid random cropping and preserve full scene context is a reasonable and underexplored direction for dense prediction. Using a lightweight RT-DETR-style hybrid encoder is sensible. The paper is also honest in places, noting that DINOv2 does better on large objects and that its pretraining cost is far lower (35 vs 800 epochs for SlotCon).

Where it falls short: the abstract says the method outperforms SOTA 'even when trained with limited annotated data and fewer fine-tuning epochs.' Table 1's most data-limited row undercuts that: at 10% labels, DINOv2 gets AP 20.1 against ProtoScale's 19.9, and DINOv2 is much better on large objects (29.58 vs 22.5). The text says ProtoScale 'consistently outperforms' on small/medium, and that is supported, but the blanket claim is not. With no error bars and a single seed, the 0.2 AP gap could be noise, but as presented the paper's own evidence fails the stated claim. This needs multi-seed runs and error bars before publication.

Eq. (2) is also sloppy. It reads as though the norm of the prototype vectors is being used as a dot product, and the softmax over N_p is not clearly aligned with the text's description of 32 learnable I_k. The baseline adaptation is under-specified: what exactly is the 'shared projection module' and what does 'integrated into our student-teacher method' mean for DINOv2 and SlotCon? If you re-trained their encoders under your framework, that is a different claim than loading official weights, and it must be described. There are no ablations on the three losses, the Gaussian prior, the auxiliary prototypes, the threshold, or prototype counts; the loss weights are asserted as 'best' with no search shown. The custom 17-class COCO+UA-DETRAC subset is fine for internal comparison but limits comparability to published numbers.

Bottom line: the method is plausible and deserves a serious referee, but the current version overreaches. A revised version with corrected equations, specified baselines, ablations, and error bars could be solid. I would send it to review rather than desk reject.

Referee Report

5 major / 6 minor

Summary. The paper proposes ProtoScale, a self-supervised representation learning method built on a student–teacher framework, which combines multi-scale semantic grouping, instance-level prototype separation, and hierarchical prototype affinities. The method is pretrained on COCO and evaluated by fine-tuning RT-DETR on a combined COCO/UA-DETRAC subset for object detection. The central claim is that ProtoScale learns object-centric representations that outperform state-of-the-art SSL baselines (DINOv1/v2, iBOT, TEC, SlotCon) under limited labeled data and few fine-tuning epochs, while using fewer parameters and avoiding the computational cost of large-scale ViT pretraining.

Significance. If the central claim is substantiated, the paper would offer a valuable compute-efficient SSL recipe for dense prediction: a lightweight 12M-parameter hybrid encoder pretrained for only 35 epochs that competes with or beats methods pretrained on ImageNet-1K or LVD-142M. The proposed ProtoScale module is a plausible and modular contribution, and the paper includes full fine-tuning epoch curves and a low-data evaluation, which are appropriate for the stated real-world deployment motivation. However, the current evidence does not fully support the headline claim. The strongest result is under 50–100% labels, where ProtoScale improves AP by 6–7 points over DINOv2; the weakest is exactly the low-data regime advertised in the abstract, where Table 1 shows DINOv2 ahead. The absence of error bars, ablations, and a precise description of baseline adaptation to RT-DETR leaves the reported gains vulnerable to implementation differences. Overall, the idea is promising and the experiments are extensive in scope, but the load-bearing comparisons need to be tightened before the central claim can be accepted.

major comments (5)

[§3.3, Eq. (2)] Equation (2) is miswritten and dimensionally unclear. It reads A_i^k = sum_{i=1}^{N_p} softmax_i( ||S_k|| * ||I_k|| / tau ) A_s^k, but the summation index i is also the index of the softmax, and the term ||S_k|| * ||I_k|| is a product of norms rather than an inner product, so the assignment weight is a scalar independent of the prototype index. The text also switches between calling S_k a 'semantic prototype' and a 'slot', and the relation to the 32 instance prototypes I_k is not formalized. Since instance grouping is one of the three core losses and a central claimed contribution, this equation must be corrected and precisely defined, including the dimensions of all tensors.
[§4.1, baseline adaptation] The fairness of the comparison depends on how each baseline backbone is integrated into RT-DETR, but this is not specified. The paper says a 'shared projection module' is attached to each encoder and that all methods are integrated into the same student–teacher pipeline, yet DINOv1/v2 and iBOT are ViT-based single-scale feature extractors while RT-DETR expects a multi-scale CNN-style pyramid. The conversion from ViT patch features to the required spatial resolutions is a non-trivial design choice and should be described concretely (e.g., projection layers, stride, feature levels) or made available in code. Without this, the reported performance gaps could reflect differences in the adaptation rather than in the SSL method itself.
[Table 1, §5.2, Abstract] The abstract and §5.2 claim that ProtoScale outperforms state-of-the-art methods 'even when trained with limited annotated data', and §5.2 states that ProtoScale 'consistently outperforms DINOv2' across data proportions. Table 1 contradicts this at the most data-limited setting: at 10% labeled data and 10 fine-tuning epochs, DINOv2 achieves AP 20.1 versus ProtoScale's 19.9, and DINOv2 is substantially better on large objects (AP_large 29.58 vs 22.5). Because experiments are single-seed and no error bars are reported, the 0.2 AP gap could be noise, but as presented the paper's own evidence does not support the 'limited annotated data' part of the headline claim. Please report multiple seeds with variance and temper the claim accordingly.
[§4.2, dataset construction] The evaluation uses a non-standard combined subset: the original class labels are 'redefined' into a 17-class taxonomy, and images are filtered to those containing at least two classes, yielding 50,022 training images. No mapping from the original COCO/UA-DETRAC classes to the 17 classes is provided, and no statistics on label distribution or filtering are given. Since all methods are fine-tuned on the same subset, the comparison may still be internally fair, but it is impossible for readers to reproduce the setup or assess whether the subset is biased toward certain object scales or categories. Please provide the class mapping, per-class instance counts, and the exact filtering procedure.
[§3.3–§3.4, experimental validation] The method introduces several hyperparameters and design choices (loss weights λ_sem=1, λ_inst=2, λ_hier=1; number of semantic and instance prototypes; Gaussian prior μ=0.5, σ=0.7; hierarchical affinity threshold 0.5; auxiliary prototypes R) but no ablation study isolates their contribution. Moreover, the loss weights are described as 'yields the best results in our experiments', indicating that they were selected using downstream performance, which is a mild form of circularity. Given that the central claim is empirical, at least a component-wise ablation and a sensitivity analysis for the most critical hyperparameters are needed to establish that the reported gains come from the ProtoScale mechanism rather than from tuning.

minor comments (6)

[§1] Typographical issues: 'row data' should be 'raw data', 'MMethods' should be 'Methods', and 'Protoscale' is inconsistently capitalized.
[§2] The related work references for TEC and [12] are incomplete: TEC is listed as NeurIPS 2022 without authors, and the entry for [12] lacks a publication venue/year in the reference list.
[§3.1] The text emphasizes avoiding disruptive cropping, but the teacher view is generated with 'random cropping, zoom out, horizontal flipping'. Please clarify what is meant by 'full scene context' given that the teacher pipeline includes cropping.
[Fig. 3] The caption says models are 'pre-trained on COCO, ImageNet-1k and LVD-142M', but the figure legend and §5.1 also report different pretraining setups (e.g., SlotCon 800 epochs on COCO). Please make the figure annotation self-contained.
[§5.1] The comparison mixes pretraining datasets and budgets (COCO 35 epochs vs ImageNet-1K 300 epochs vs LVD-142M 300 epochs). This is acceptable if the claim is about total compute efficiency, but the compute/epoch counts should be stated clearly in the figure or table so readers can interpret the comparison.
[General] No code, configuration files, or trained model checkpoints are provided. For a method whose main contribution is empirical, releasing code and exact training configurations would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central claim is an empirical benchmark comparison, not a derivation from fitted inputs; the Table 1 inconsistency is a correctness concern, not circularity.

full rationale

The central claim of the paper is empirical: ProtoScale pretraining is evaluated by fine-tuning the resulting backbone for object detection on a combined COCO/UA-DETRAC subset and comparing COCO AP against several external SSL baselines. The pretraining objective is composed of three KL-divergence terms between student and teacher attention maps (Eqs. 1–2 plus the hierarchical attention), and the downstream evaluation metric is a standard detection AP. No equation defines the reported AP in terms of the pretraining loss or in terms of fitted parameters. The learnable prototypes and the loss weights (lambda_sem, lambda_inst, lambda_hier) = (1,2,1) are stated as experimental configuration choices, not as quantities derived from the downstream benchmark. The comparison is self-contained in the sense that it relies on externally pre-trained checkpoints (DINOv1/DINOv2/iBOT/TEC/SlotCon) or on the paper's own pretraining, and no load-bearing argument is justified by self-citation: none of the cited references are prior works by the present authors. The paper also explicitly acknowledges one limitation in Section 5.1: DINOv2 slightly outperforms ProtoScale on large objects. There is a genuine internal inconsistency: Table 1 shows DINOv2 with AP 20.1 vs ProtoScale 19.9 at 10% labeled data and 10 fine-tuning epochs, while Section 5.2 claims ProtoScale 'consistently outperforms DINOv2' across data proportions. This is a correctness/reproducibility issue (single-seed, no error bars), not a circularity; it does not mean the result reduces to its inputs by construction. Therefore no circular step is identifiable, and the appropriate score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 4 invented entities

The central claim depends on several hand-chosen hyperparameters and architectural assumptions. No free parameter is fitted with uncertainty quantification, and none of the invented grouping components is independently validated.

free parameters (5)

loss weights (λ_sem, λ_inst, λ_hier) = (1, 2, 1)
Selected as the best in experiments (Section 3.4); tuned based on downstream validation.
number of semantic prototypes = 256
Ad hoc architectural choice; no justification or ablation.
number of instance prototypes = 32
Ad hoc architectural choice; no justification or ablation.
Gaussian prior parameters (μ, σ) = (0.5, 0.7)
Hand-chosen to reduce collapse (Section 3.3); values are not derived.
hierarchical affinity threshold = 0.5
Hand-chosen to merge instances (Section 3.3); no sensitivity analysis.

assumptions (5)

domain assumption A student-teacher setup with EMA updates yields stable and useful self-supervised targets.
Invoked in Section 3.4 without proof; standard in SSL but assumption that it transfers to dense grouping.
domain assumption KL divergence between student and teacher attention maps enforces consistent grouping.
Used in the loss (Section 3.4); no evidence given that attention-map matching preserves object structure.
ad hoc to paper Gaussian centering of attention maps prevents collapse.
Introduced in Section 3.3 with parameters chosen by hand; not derived or validated.
domain assumption The RT-DETR hybrid encoder's multi-scale features are suitable for prototype grouping at three scales.
Borrowed from RT-DETR (Section 3.2); assumed to work without modification.
domain assumption Images with at least two classes form a representative evaluation set for object detection.
Dataset construction in Section 4.2; this assumption affects the validity of downstream comparisons.

invented entities (4)

Semantic prototypes S_k
purpose: Cluster spatial features into semantic groups at each scale.
Internal learned parameters; no external falsifiable prediction.
Instance prototypes I_k
purpose: Separate object instances across the batch.
Internal learned parameters; no external validation such as instance segmentation benchmarks.
Auxiliary prototypes R
purpose: Absorb noisy activations to stabilize training.
Borrowed conceptually from registers (ref [12]) but repurposed; no ablation.
Hierarchical affinity matrix H_k
purpose: Model object-part relationships between instances.
Predicted by an MLP on instance prototypes; no direct evidence it captures true hierarchy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Self-supervised structured object representation learning." pith.science (2026). https://pith.science/paper/6GBK5GSI

@misc{pith2026250819864,
  author       = {Pith},
  title        = {Pith review of: Self-supervised structured object representation learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6GBK5GSI}},
  note         = {Machine review of arXiv:2508.19864}
}

read the original abstract

Self-supervised learning (SSL) has emerged as a powerful technique for learning visual representations. While recent SSL approaches achieve strong results in global image understanding, they are limited in capturing the structured representation in scenes. In this work, we propose a self-supervised approach that progressively builds structured visual representations by combining semantic grouping, instance level separation, and hierarchical structuring. Our approach, based on a novel ProtoScale module, captures visual elements across multiple spatial scales. Unlike common strategies like DINO that rely on random cropping and global embeddings, we preserve full scene context across augmented views to improve performance in dense prediction tasks. We validate our method on downstream object detection tasks using a combined subset of multiple datasets (COCO and UA-DETRAC). Experimental results show that our method learns object centric representations that enhance supervised object detection and outperform the state-of-the-art methods, even when trained with limited annotated data and fewer fine-tuning epochs.

Figures

Figures reproduced from arXiv: 2508.19864 by the authors.

**Figure 1.** Overview of the proposed training framework integrating multi-view learning, multi-scale feature encoding, and the ProtoScale grouping module. The teacher network (depicted in blue) is frozen (SG: stop-gradient), while the student network (depicted in red) is trainable. globally augmented image obtained through a sequence of spatial transformations, including random cropping, zoom out, horizontal flipping, and phot… view at source ↗

**Figure 2.** Multi-level grouping process in ProtoScale module. 3.3 ProtoScale: a multi-level grouping modules Once the features are extracted from the encoders, they are processed by our proposed grouping module, ProtoScale ( [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Main Transfer Results with SSL Pre-training over Epochs 5 Results and discussion In this section, we evaluate the performance of our SSL model through a downstream object detection task. We compare ProtoScale with several SOTA SSL methods on a combined subset of multiple datasets. Performance is reported using the standard COCO detection metrics. AP and AR denote the mean Average Precision and mean Average Recall, … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 23 canonical work pages

[1]

In: International Conference on Learning Representations (ICLR) (2021)

Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-traini ng of image transformers. In: International Conference on Learning Representations (ICLR) (2021)

work page 2021
[2]

In: Comput er Vision – ECCV 2020, 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I

Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillo v, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Comput er Vision – ECCV 2020, 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I. Lecture Notes in Computer Science, vol. 12346, pp. 213–22 9 (2020)

work page 2020
[3]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICC V)

Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., B ojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transforme rs. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICC V). pp. 9915–9925 (October 2021)

work page 2021
[4]

In: III, H.D

Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations. In: III, H.D. , Singh, A. (eds.) Proceed- ings of the 37th International Conference on Machine Learni ng. Proceedings of Machine Learning Research, vol. 119, pp. 1597–1607. PMLR (J uly 2020)

work page 2020
[5]

IEEE Conference on Computer V ision and Pattern Recognition (CVPR) pp

Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L .: Imagenet: A large-scale hierarchical image database. IEEE Conference on Computer V ision and Pattern Recognition (CVPR) pp. 248–255 (2009)

work page 2009
[6]

In: Pr oceedings of the 2019 Conference of the North American Chapter of the ACL (NAACL-H LT)

Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre -training of deep bidi- rectional transformers for language understanding. In: Pr oceedings of the 2019 Conference of the North American Chapter of the ACL (NAACL-H LT). pp. 4171– 4186 (2019)

work page 2019
[7]

In: International Conference on Learning Representations (20 21)

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn , D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszk oreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recogn ition at scale. In: International Conference on Learning Representations (20 21)

work page
[8]

In: Proceedings of the 36th Conference on Neural Inform ation Processing Sys- tems (NeurIPS) (2022) 12 O

Gao, S., Zhou, P., Cheng, M.M., Yan, S.: Towards sustainab le self-supervised learn- ing. In: Proceedings of the 36th Conference on Neural Inform ation Processing Sys- tems (NeurIPS) (2022) 12 O. Hadjerci et al

work page 2022

Show all 24 references

[9]

In: Proceedings of the IEE E/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum con trast for unsupervised visual representation learning. In: Proceedings of the IEE E/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9726–9 735 (June 2020)

2020
[10]

In: Proceedings of the IEEE Conference on Computer Vision and Pa ttern Recognition (CVPR)

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learnin g for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pa ttern Recognition (CVPR). pp. 770–778 (2016)

2016
[11]

, Vinyals, O., Carreira, J.: Eﬃcient visual pretraining with contrastive detection

Hénaﬀ, O.J., Koppula, S., Alayrac, J.B., van den Oord, A. , Vinyals, O., Carreira, J.: Eﬃcient visual pretraining with contrastive detection . In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICC V). pp. 10086– 10096 (October 2021)

2021
[12]

Jiang, N., Dravid, A., Efros, A., Gandelsman, Y.: Vision transformers don’t need trained registers (2025)

2025
[13]

In: European Conference on Computer Vision (ECCV)

Lin, T., Maire, M., Belongie, S.J., Bourdev, L.D., Girsh ick, R.B., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European Conference on Computer Vision (ECCV) . Lecture Notes in Computer Science, vol. 8693, pp. 74...

2014
[14]

In: Advances in Neural Information Processing System s (NeurIPS)

Locatello, F., Weissenborn, D., Unterthiner, T., Mahen dran, A., Heigold, G., Uszkoreit, J., Dosovitskiy, A., Kipf, T.: Object-centric l earning with slot atten- tion. In: Advances in Neural Information Processing System s (NeurIPS). vol. 33, pp. 11528–11539 (2020)

2020
[15]

In: Proceedings of the IEEE/ CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024)

Oquab, M., Darcet, T., Moutakanni, T., Ramé, A., Mazur, D ., Genet, A., Mann, E., Misra, I., Neverova, N., Labatut, P., Jegou, H.: Dinov2: Learning robust visual features without supervision. In: Proceedings of the IEEE/ CVF Conference on Computer Vision and Pattern Recognitio...

2024
[16]

OpenAI Technical Report (2018)

Radford, A., Narasimhan, K., Salimans, T., Sutskever, I .: Improving language un- derstanding by generative pre-training. OpenAI Technical Report (2018)

2018
[17]

arXiv preprint arXiv:2304.05387 (2023)

Rambhatla, S.S., Misra, I., Chellappa, R., Shrivastava , A.: Multiple object lo- calization with self-supervised transformers for object d iscovery. arXiv preprint arXiv:2304.05387 (2023)

2023 arXiv
[18]

, Joulin, A., Grave, E., Lample, G.: Llama: Open and eﬃcient foundation language models

Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lach aux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A. , Joulin, A., Grave, E., Lample, G.: Llama: Open and eﬃcient foundation language models. CoRR, abs/2302.13971 (2023)

2023 arXiv
[19]

In: Proceedings of th e IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Wang, X., Girdhar, R., Yu, S.X., Misra, I.: Cut and learn f or unsupervised object detection and instance segmentation. In: Proceedings of th e IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3124 –3134 (June 2023)

2023
[20]

In: Proceedings of the 35th International Confer- ence on Neural Information Processing Systems (NeurIPS) (2 021)

Wei, F., Gao, Y., Wu, Z., Hu, H., Lin, S.: Aligning pretrai ning for detection via object-level contrastive learning. In: Proceedings of the 35th International Confer- ence on Neural Information Processing Systems (NeurIPS) (2 021)
[21]

Computer Vision and Image Understanding 193, 102907 (2020)

Wen, L., Du, D., Cai, Z., Lei, Z., Chang, M.C., Qi, H., Lim, J., Yang, M.H., Lyu, S.: Ua-detrac: A new benchmark and protocol for multi-object de tection and tracking. Computer Vision and Image Understanding 193, 102907 (2020)

2020
[22]

In: Advances in Neural Information Processing Systems

Wen, X., Zhao, B., Zheng, A., Zhang, X., Qi, X.: Self-supe rvised visual repre- sentation learning with semantic grouping. In: Advances in Neural Information Processing Systems. vol. 35, pp. 22644–22658 (2022)

2022
[23]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Zhao, Y., Lv, W., Xu, S., Wei, J., Wang, G., Dang, Q., Liu, Y ., Chen, J.: Detrs beat yolos on real-time object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. xxxx –xxxx (June 2024)

2024
[24]

In: Proceedings of the 10th International Conference on Learning Representations (ICLR) (2022)

Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: Image bert pre- training with online tokenizer. In: Proceedings of the 10th International Conference on Learning Representations (ICLR) (2022)

2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

In: International Conference on Learning Representations (ICLR) (2021)

Bao, H., Dong, L., Piao, S., Wei, F.: Beit: Bert pre-traini ng of image transformers. In: International Conference on Learning Representations (ICLR) (2021)

work page 2021

[2] [2]

In: Comput er Vision – ECCV 2020, 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I

Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillo v, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Comput er Vision – ECCV 2020, 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I. Lecture Notes in Computer Science, vol. 12346, pp. 213–22 9 (2020)

work page 2020

[3] [3]

In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICC V)

Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., B ojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transforme rs. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICC V). pp. 9915–9925 (October 2021)

work page 2021

[4] [4]

In: III, H.D

Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations. In: III, H.D. , Singh, A. (eds.) Proceed- ings of the 37th International Conference on Machine Learni ng. Proceedings of Machine Learning Research, vol. 119, pp. 1597–1607. PMLR (J uly 2020)

work page 2020

[5] [5]

IEEE Conference on Computer V ision and Pattern Recognition (CVPR) pp

Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L .: Imagenet: A large-scale hierarchical image database. IEEE Conference on Computer V ision and Pattern Recognition (CVPR) pp. 248–255 (2009)

work page 2009

[6] [6]

In: Pr oceedings of the 2019 Conference of the North American Chapter of the ACL (NAACL-H LT)

Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre -training of deep bidi- rectional transformers for language understanding. In: Pr oceedings of the 2019 Conference of the North American Chapter of the ACL (NAACL-H LT). pp. 4171– 4186 (2019)

work page 2019

[7] [7]

In: International Conference on Learning Representations (20 21)

Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn , D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszk oreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recogn ition at scale. In: International Conference on Learning Representations (20 21)

work page

[8] [8]

In: Proceedings of the 36th Conference on Neural Inform ation Processing Sys- tems (NeurIPS) (2022) 12 O

Gao, S., Zhou, P., Cheng, M.M., Yan, S.: Towards sustainab le self-supervised learn- ing. In: Proceedings of the 36th Conference on Neural Inform ation Processing Sys- tems (NeurIPS) (2022) 12 O. Hadjerci et al

work page 2022