Pith. sign in

REVIEW 3 major objections 6 minor 96 references

ViT-Split: Unleashing the Power of Vision Foundation Models via Efficient Splitting Heads

T0 review · 3 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 Frozen vision models can match full adapters with two small heads

desk verdict Useful and mostly honest method paper: the split idea works for segmentation but the efficiency claim does not transfer to detection, and the paper admits as much in the appendix. read the letter →

arxiv 2506.03433 v2 pith:Q2BQZ5JN submitted 2025-06-03 cs.CV

Yifan Li , Xin Li , Tianqin Li , Wenbin He , Yu Kong , Liu Ren This is my paper
classification cs.CV
keywords visionfoundationmodelsparameter-efficientfine-tuningsemanticsegmentationobjectdetectionmonoculardepthestimationvisualquestionansweringDINOv2layersplit
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 tries to establish that a vision foundation model does not needed to be fine-tuned, or hooked to a CNN branch, to adapt to dense prediction tasks. It claims that the layers of models like DINOv2 split naturally into an early feature extractor and a late task-specific adapter, so the late layers can be copied out, frozen-backbone features can be tapped at multiple scales, and only two small heads need training. On ADE20K segmentation this reaches comparable or better mIoU than existing VFM adapters while tuning about a fifth to a quarter of the parameters and training up to 4x faster. If this holds, it changes what an adapter for a vision foundation model needs to be.

What carries the argument

The load-bearing mechanism is the layer split itself, revealed by CKA similarity: early VFM layers act as a generic feature extractor and late layers act as a task-specific adapter. From that split the architecture follows: features from layer $L - K_t$ feed the task head $g_{\theta_t}$, a replica of the last $K_t$ transformer layers; $K_p$ frozen layer features chosen by uniform sampling with indices $b + \mathrm{round}(i \cdot \delta)$, $\delta = (L - b - 1)/(K_p - 1)$, are concatenated and passed through the prior head $g_{\theta_p}$, a $1 \times 1$ convolution followed by a $3 \times 3$ deformable convolution; and the fusion net $g_{\theta_f}$ concatenates the task and prior feature maps. The mechanism's work is to let gradients reach only the small heads, never the early backbone, while multi-scale frozen features stand in for the CNN branch that other adapters need.

What would settle it

Train ViT-Split with the task head alone and then unfreeze the early backbone layers; if unfreezing early layers substantially raises validation mIoU, the claimed extractor/adapter split is not doing the work.

Watch

Extended reading notes

Core claim

The paper's central claim is that the last few layers of a frozen VFM, when replicated as a task-specific head and trained alone, carry out task adaptation as effectively as fine-tuning the whole backbone, and that a separate prior head aggregating multi-scale frozen features supplies the low-level and semantic knowledge that previous adapters obtained from a CNN branch. Evidence for the split comes from CKA similarity matrices and feature visualizations showing that early-layer features stay similar across pretraining, segmentation, and detection, while late-layer features diverge by task. ViT-Split therefore removes the CNN branch, freezes the entire backbone, and trains only a task head initialized from the last few layers, a two-layer CNN prior head over uniformly sampled frozen layers, and a small fusion net. The paper reports that this setup matches or beats ViT-Adapter and ViT-CoMer on ADE20K, Cityscapes, COCO detection, NYU-V2 depth, and several VQA benchmarks while tuning a small fraction of the parameters.

Load-bearing premise

The argument stands or falls on the premise that the last few layers of a frozen VFM, copied out and trained alone, can perform task adaptation just as well as fine-tuning the entire backbone.

Editorial extensions

If this is right

  • A large segmentation head such as Mask2Former or UperNet is not required to exploit a VFM; a linear head plus the two ViT-Split heads reaches comparable mIoU with far fewer tuned parameters.
  • One frozen backbone can serve many tasks at once, because only task-specific task heads, prior heads, and fusion nets need to be stored and trained.
  • Training cost scales with the size of the heads rather than the depth of the backbone, so the speed advantage should grow as VFMs get larger.
  • Detection is the hard case: because DINOv2's pretraining does not align with detection targets, ViT-Split needs many more tuned layers there, implying the layer split has task-dependent limits.
  • Longer training pushes ViT-Split further, with 160K iterations raising ADE20K small-model mIoU to 52.2%, above the fine-tuned UperNet baseline.

Reading between the lines

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

  • Beyond the paper: if the layer split is a general property of self-supervised transformers, the same copy-the-last-layers trick could replace per-layer adapters or prompt tokens in other modalities, not just vision.
  • Beyond the paper: the paper does not isolate whether initialization matters; comparing a task head initialized from copied weights against a randomly initialized task head would tell whether the copied weights carry the gain or just provide a convenient architecture.
  • Beyond the paper: the sparse gate's near-identical accuracy to uniform sampling on segmentation suggests that a hyperparameter-free version of ViT-Split is within reach, with the layer selection learned instead of hand-set.
  • Beyond the paper: because ViT-Split leaves the VFM's features unaltered, it may combine with prompt tuning or low-rank weight tuning rather than compete with them, potentially stacking efficiency mechanisms.
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

3 major / 6 minor

Summary. The paper proposes ViT-Split, a parameter- and time-efficient adapter for vision foundation models. The central idea is to freeze the VFM backbone and append two lightweight trainable modules: a task head, initialized from the last Kt layers of the same VFM, and a prior head that aggregates Kp multi-scale features from the frozen backbone, followed by a CNN fusion net. The method is motivated by CKA and feature-visualization evidence that DINOv2 layers can be divided into an early feature extractor and a late task-specific adapter. Experiments cover semantic segmentation on ADE20K, Cityscapes, and Pascal Context; object detection and instance segmentation on COCO; monocular depth estimation on NYU-V2; and visual question answering with LLaVA-1.5. The paper reports competitive or better accuracy than prior VFM adapters while tuning far fewer parameters and reducing training time, with the 4x speedup claim made for ADE20K segmentation.

Significance. If the claims hold, ViT-Split offers a practically valuable recipe: a frozen DINOv2 backbone plus two small heads can compete with full fine-tuning and with CNN-augmented VFM adapters on dense prediction tasks. The segmentation results are strong and the paper includes extensive ablations, hyperparameter sensitivity analyses, an appendix-level fairer comparison (Table 14), and results on multiple VFMs and tasks. The paper also releases code, which supports reproducibility. However, the significance is tempered by the detection results: on COCO the required Kt is so large that the task head is nearly the whole backbone, reducing the claimed efficiency gain to a roughly 4% parameter saving and a 42% training-time reduction. Because the paper's framing in the title and abstract generalizes the extractor/adapter split to 'various tasks', the actual scope is narrower than advertised unless the task-dependence is explicitly quantified and explained in the main text.

major comments (3)
  1. [Sec. 4.2 and Appendix A.1, Table 10; Appendix D] The detection experiments directly limit the paper's central efficiency claim and should be confronted in the abstract and introduction. Table 10 shows Kt = 23/24 for ViT-Split-L on COCO, so the 'task head' is initialized from almost the entire backbone, and Table 5 reports 348M trainable parameters versus 363M for ViT-CoMer-L, with the main text stating a 42% training-time reduction rather than the multi-fold speedup advertised elsewhere. Appendix D explicitly acknowledges that DINOv2 features are weakly aligned with detection and that a larger Kt is therefore necessary. Since the title and abstract claim general efficiency across 'various tasks', the paper should either restrict the efficiency claim to tasks where the prior aligns well (dense prediction, as on ADE20K) or provide a quantitative recipe, e.g., reporting the trainable-parameter fraction and wall-clock speedup for every benchmark and stating the condition under which the 4x figure holds.
  2. [Tables 1, 5, and Appendix Table 14] The headline comparisons use different downstream heads and training schedules for ViT-Split versus the baselines. In Table 1, ViT-Split uses a linear head trained for 40k iterations while ViT-Adapter and ViT-CoMer use UperNet heads trained for 160k iterations; Table 5 similarly compares Mask R-CNN results under different schedules and head configurations. Appendix Table 14 removes the schedule confound by training all methods for 40k iterations, but still couples UperNet with the baselines and a linear head with ViT-Split. Because a larger head can change both accuracy and wall-clock time, the claim that the architecture itself, rather than the choice of head, drives the gains would be strengthened by at least one matched-head experiment, for example feeding ViT-Split features into the same UperNet or Mask R-CNN head used for the baselines, or training the baselines with a linear head.
  3. [Tables 1, 4, and 5] All reported results appear to come from single runs, with no error bars or repeated-seed statistics. Some of the comparisons that support the 'comparable or better' claim are within a few tenths of a point (e.g., Table 5: ViT-Split-B 51.8 vs. ViT-CoMer-B 52.0 APbbox; Table 4: LLaVA-1.5 + ViT-Split 78.2 vs. 78.5 on VQAv2). Without variance estimates or at least two seeds, the reader cannot tell whether these differences are significant. Since the central claims include both 'better' and 'comparable' outcomes, the absence of run-to-run variability is a load-bearing gap, especially for the detection and VQA tables.
minor comments (6)
  1. [Appendix A.1, Table 10] The table heading 'Cityscales' is a typo and should read 'Cityscapes'; the same typo appears in the main text before Table 3.
  2. [Appendix B.1, Eq. (5)] The straight-through estimator equation is notationally ambiguous: the same symbol Gsp appears on both sides. Please write the detached term explicitly, e.g., Gsp = G_hard + G - stop_gradient(G), so the forward and backward paths are clear.
  3. [Sec. 4.4, Table 6] The table caption should state directly which row is the frozen final-layer baseline and which row is full fine-tuning; the current explanation in the text is easy to misread, especially for readers who only inspect the table.
  4. [Fig. 8 and Appendix C] The speedup numbers are inconsistent: the text says about 4x for the small model and 3x for the base model, while Appendix C states 2.4-5x training speedup and 1.4-3x speedup over full fine-tuning. Please harmonize these statements and specify whether they refer to per-iteration wall-clock time, total training time, or both.
  5. [Table 2] The entry for ViT-Adapter-G* says it is 'implemented without tuning the whole backbone', but the table does not specify which components were trained (adapter only? adapter plus head?) or the exact iteration count. Please clarify in the table caption or in Sec. 4.1.
  6. [Sec. 4.3 and Appendix A.3] The VQA adaptation uses Kt = 1 and Kp = 1, so only one layer of CLIP is tuned and only the last-layer feature is used as prior. This is a much weaker instantiation of the extractor/adapter split than in the dense-prediction experiments; a sentence explaining why this setting was chosen and whether it is still representative of ViT-Split would help the reader.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the extractor/adapter split is an empirical observation, and all headline numbers are measured against external benchmarks with honest scope caveats.

full rationale

The paper's central claim is that freezing a VFM and tuning a task head (copied from the last Kt layers) plus a prior head over multi-scale frozen features matches or beats VFM adapters at lower training cost. This claim is supported by external benchmark results (ADE20K, Cityscapes, COCO, NYU-V2, VQA) and by ablations (Tables 6-9), not by a fitted parameter masquerading as a prediction. The extractor/adapter distinction (Sec. 3.1) is justified by CKA and feature visualizations; it motivates the architecture but is not derived from the benchmark numbers. Kt and Kp are hyperparameters tuned per task (Table 10, Fig. 11); reporting their chosen values is standard model selection, not circularity. The only overlapping-author reference, [50], is a VLLM survey cited in the VQA setup and carries no load-bearing argument. The paper's own limitations weaken the scope of the efficiency claim: Appendix D states that DINOv2 features are weakly aligned with detection, requiring Kt=23/24 for ViT-Split-L on COCO (Table 10), so detection saves only about 4% of parameters and roughly 42% of training time, and Appendix F limits demonstrated VFMs to DINOv2 and CLIP. These are honest scope caveats, not circular reductions. No equation or result in the paper is equivalent by construction to its inputs.

Assumptions & free parameters 4 free parameters · 6 assumptions · 3 invented entities

The method relies on a small set of tunable hyperparameters (Kt, Kp, b, and the task-head LR scale) and on the qualitative layer-split assumption. The two heads and the sparse gate are invented architectural components with only within-paper validation. The axioms are mostly domain assumptions about how VFM representations behave and how the components should be designed. No new physical or mathematical entities are introduced.

free parameters (4)
  • Kt (number of copied last layers for task head) = 3 to 23 depending on task and backbone size
    Chosen per task and model size (Table 10, Fig. 11). Segmentation uses Kt=3-4 for small/base, detection uses up to Kt=23, VQA uses Kt=1. The paper reports sensitivity analysis but treats Kt as a tunable hyperparameter.
  • Kp (number of sampled prior layers) = 1 to 14 depending on task
    Kp=4 works for most tasks, but VQA uses Kp=1 and DINOv2-G uses Kp=14 (Table 10, A.1). This is a free parameter fitted to validation performance.
  • b (starting layer index for uniform sampling) = 2, 3, 6, 23, or 26
    Set to skip noisy early layers; varies by task and model size (Table 10, Eq. 2). It is a manually chosen offset.
  • task-head learning rate scale = 0.1
    The learning rate for the task head is reduced by a factor of 0.1 relative to the other components (Sec. 4.1). This constant is chosen without a reported sensitivity study.
assumptions (6)
  • standard math Centered Kernel Alignment (CKA) reliably measures layer-feature similarity for the split observation (Sec. 3.1, [44]).
    The paper uses CKA plots to argue that early layers are similar to each other and late layers are similar to each other. This is a standard tool, but the threshold for 'similar' is qualitative.
  • domain assumption VFMs like DINOv2 can be partitioned into an early-layer extractor and a late-layer task-specific adapter (Sec. 3.1, Figs. 3-4).
    This is the core observation motivating the architecture. It is based on feature visualization and CKA, and the paper acknowledges it holds for 'several VFMs' but not all.
  • ad hoc to paper A task head copied from the last Kt layers, trained alone, can adapt the representation as effectively as fine-tuning the entire backbone.
    The paper demonstrates this empirically (Tab. 6 shows task head alone reaches 49.5 vs 53.8 for full fine-tuning on small, and with prior head reaches 51.6) but offers no proof. It is the load-bearing premise of the efficiency gain.
  • domain assumption Frozen intermediate features from uniformly sampled layers remain useful after task-head training and can be aggregated by a small CNN.
    The prior head is designed on this premise. The ablation in Tab. 6 supports it on ADE20K, but the mechanism is not established and the utility may vary across tasks.
  • domain assumption A 1x1 convolution followed by a 3x3 deformable convolution provides sufficient capacity to aggregate multi-scale frozen features.
    This is a specific architectural choice for the prior head and fusion net. It is validated only by the single ablation in Tab. 6, with no formal or broader justification.
  • standard math Straight-Through Estimator approximates gradients for the non-differentiable top-K sparse gate (Appendix B.1, Eq. 5).
    STE is a standard technique for discrete selection. The paper uses it for the sparse-gate variant, which is only evaluated on segmentation and performs comparably to uniform sampling.
invented entities (3)
  • Task head (copied last layers of the VFM)
    purpose: Learn task-specific features while avoiding gradient propagation through the backbone
    Its value is shown only by ablation within the paper (Tab. 6); no external falsifiable handle is provided.
  • Prior head (two-layer CNN aggregating multi-scale frozen features)
    purpose: Aggregate multi-scale prior features from the frozen VFM to reduce overfitting and add local representation
    Ablation Tab. 6 shows it improves mIoU, but only on benchmarks in the same paper; no outside prediction.
  • Sparse gate (learnable layer-selection mechanism)
    purpose: Select top-Kp layer features with STE in a differentiable way
    Appendix B.1; segmentation experiments show it matches uniform sampling, but no independent evidence exists.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ViT-Split: Unleashing the Power of Vision Foundation Models via Efficient Splitting Heads." pith.science (2026). https://pith.science/paper/Q2BQZ5JN

@misc{pith2026250603433,
  author       = {Pith},
  title        = {Pith review of: ViT-Split: Unleashing the Power of Vision Foundation Models via Efficient Splitting Heads},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q2BQZ5JN}},
  note         = {Machine review of arXiv:2506.03433}
}
abstract

Vision foundation models (VFMs) have demonstrated remarkable performance across a wide range of downstream tasks. While several VFM adapters have shown promising results by leveraging the prior knowledge of VFMs, we identify two inefficiencies in these approaches. First, the interaction between convolutional neural network (CNN) and VFM backbone triggers early layer gradient backpropagation. Second, existing methods require tuning all components, adding complexity. Besides, these adapters alter VFM features, underutilizing the prior knowledge. To tackle these challenges, we propose a new approach called ViT-Split, based on a key observation: the layers of several VFMs, like DINOv2, can be divided into two distinct components: an extractor for learning low-level features and an adapter for learning task-specific features. Leveraging this insight, we eliminate the CNN branch and introduce two heads, task head and prior head, to the frozen VFM. The task head is designed to learn task-specific features, mitigating the early gradient propagation issue. The prior head is used to leverage the multi-scale prior features from the frozen VFM, reducing tuning parameters and overfitting. Extensive experiments on various tasks (e.g., segmentation, detection, depth estimation, and visual question answering) validate the effectiveness and efficiency of ViT-Split. Specifically, ViT-Split reduces training time up to $4\times$ while achieving comparable or even better results on ADE20K, compared to other VFM adapters.

Figures

Figures reproduced from arXiv: 2506.03433 by the authors.

Figure 1
Figure 1. Comparison between previous VFM adapters and ours. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison with previous VFM adapters (ViT-Adapter [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The CKA comparison of layer features across different [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Comparison of DINOv2-S layer features across different [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]
Figure 5
Figure 5. Figure 5: The framework of ViT-Split. ViT-Split introduces two [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: The illustration of the CNN fusion architecture. It is [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Inference comparison: (a) Previous VFM adapters vs. (b) Our ViT-Split. ViT-Split is efficient during inference for multiple tasks. 0 2 4 6 8 GFLOPs Time Params 6.1 3.5 5.7 6.7 5.0 6.1 1.0 1.0 1.0 ViT-S on ADE20K ViT-Adapter ViT-CoMer ViT-Split (a) 0 2 4 GFLOPs Time Par…
Figure 8
Figure 8. Figure 8: Comparison of time complexity for VFM adapters on [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 9
Figure 9. Figure 9: Segmentation results and parameters on ADE20K with [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Illustration of our proposed layer selection methods: [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11: Parameter sensitivity analysis of Kt and Kp in ViT-Split. The experiments are conducted using ViT-Split-S on ADE20K. on ADE20K, using a UperNet head for the baselines and a linear head for ViT-Split. For VFM-adapters, we adopt a learning rate schedule similar to that …
Figure 12
Figure 12. Figure 12: The CKA of SAM (a) and MAE (b). (c) Training com [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: Further comparison of DINOv2-S layer features across original features, segmentation, and detection tasks. In each figure, [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 14
Figure 14. Figure 14: Semantic segmentation and instance segmentation results based on our ViT-Split-L ( [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: The CKA visualizations of different sizes of DINOv2. [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

96 extracted references · 61 canonical work pages

  1. [1]

    Attention attention everywhere: Monocular depth prediction with skip attention

    Ashutosh Agarwal and Chetan Arora. Attention attention everywhere: Monocular depth prediction with skip attention. In WACV, pages 5861–5870, 2023. 2

    2023

  2. [2]

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

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bo- janowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In CVPR, pages 15619–15629, 2023. 1, 3

    2023

  3. [3]

    Foundational models defining a new era in vision: A survey and outlook

    Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook. arXiv preprint arXiv:2307.13721, 2023. 2

    arXiv 2023

  4. [4]

    Qwen-vl: A frontier large vision-language model with versatile abilities

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 3, 7

    arXiv 2023

  5. [5]

    Beit: Bert pre-training of image transformers

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2022. 1, 3

    2022

  6. [6]

    Adabins: Depth estimation using adaptive bins

    Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, pages 4009–4018, 2021. 2

    2021

  7. [7]

    Language models are few-shot learners

    Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3

    arXiv 2005

  8. [8]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In ICCV, pages 9650–9660, 2021. 1

    2021

Show all 96 references
  1. [9]

    Mmdetection: Open mmlab detection tool- box and benchmark

    Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection tool- box and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7

    1906 arXiv

  2. [10]

    Shikra: Unleashing multi- modal llm’s referential dialogue magic

    Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multi- modal llm’s referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 7

    2023 arXiv

  3. [11]

    Mixformer: Mixing features across windows and dimensions

    Qiang Chen, Qiman Wu, Jian Wang, Qinghao Hu, Tao Hu, Errui Ding, Jian Cheng, and Jingdong Wang. Mixformer: Mixing features across windows and dimensions. In CVPR, pages 5249–5259, 2022. 7

    2022

  4. [12]

    Adaptformer: Adapt- ing vision transformers for scalable visual recognition

    Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapt- ing vision transformers for scalable visual recognition. In NeurIPS, pages 16664–16678, 2022. 2, 3

    2022

  5. [13]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge- offrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597–1607, 2020. 1, 3

    2020

  6. [14]

    Vision transformer adapter for dense predictions

    Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. In ICLR, 2023. 1, 2, 3, 5, 6, 7

    2023

  7. [15]

    Internvl: Scaling up vision foundation mod- els and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation mod- els and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 3

    2024

  8. [16]

    Masked-attention mask transformer for universal image segmentation

    Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexan- der Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 2, 6

    2022

  9. [17]

    Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality. See https://vicuna. lmsys. org (accessed 14 April 2023) , 2(3):6,

    2023

  10. [18]

    Twins: Revisiting the design of spatial attention in vision transformers

    Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haib- ing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In NeurIPS, pages 9355–9366, 2021. 5, 7

    2021

  11. [19]

    Mmsegmentation: Open- mmlab semantic segmentation toolbox and benchmark,

    MMSegmentation Contributors. Mmsegmentation: Open- mmlab semantic segmentation toolbox and benchmark,

  12. [20]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213–3223, 2016. 5

    2016

  13. [21]

    Deformable convolutional networks

    Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In ICCV, pages 764–773, 2017. 4

    2017

  14. [22]

    Instructblip: Towards general- purpose vision-language models with instruction tuning

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: Towards general- purpose vision-language models with instruction tuning. In NeurIPS, 2023. 7

    2023

  15. [23]

    Scaling vision transformers to 22 billion pa- rameters

    Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdul- mohsin, et al. Scaling vision transformers to 22 billion pa- rameters. In ICML, pages 7480–7512, 2023. 3

    2023

  16. [24]

    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, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. In ICLR, 2020. 2

    2020

  17. [25]

    Eva: Exploring the limits of masked visual representa- tion learning at scale

    Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representa- tion learning at scale. In CVPR, pages 19358–19369, 2023. 1, 3, 4

    2023

  18. [26]

    Deep ordinal regression net- work for monocular depth estimation

    Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat- manghelich, and Dacheng Tao. Deep ordinal regression net- work for monocular depth estimation. InCVPR, pages 2002– 2011, 2018. 2

    2002

  19. [27]

    Making the v in vqa matter: Elevating the role of image understanding in visual question answer- ing

    Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Ba- tra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answer- ing. In CVPR, pages 6904–6913, 2017. 7

    2017

  20. [28]

    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 Ghesh- laghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. InNeurIPS, pages...

  21. [29]

    A survey on self-supervised learning: Algorithms, applications, and future trends

    Jie Gui, Tuo Chen, Jing Zhang, Qiong Cao, Zhenan Sun, Hao Luo, and Dacheng Tao. A survey on self-supervised learning: Algorithms, applications, and future trends. IEEE TPAMI, 2024. 1

    2024

  22. [30]

    Vizwiz grand challenge: Answering visual questions from blind people

    Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In CVPR, pages 3608–3617, 2018. 7

    2018

  23. [31]

    Flatten transformer: Vision transformer using fo- cused linear attention

    Dongchen Han, Xuran Pan, Yizeng Han, Shiji Song, and Gao Huang. Flatten transformer: Vision transformer using fo- cused linear attention. In ICCV, 2023. 7

    2023

  24. [32]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016. 2

    2016

  25. [33]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Gir- shick. Mask r-cnn. In ICCV, pages 2961–2969, 2017. 3, 7

    2017

  26. [34]

    Momentum contrast for unsupervised visual rep- resentation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual rep- resentation learning. In CVPR, pages 9729–9738, 2020. 1, 3

    2020

  27. [35]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000–16009, 2022. 1, 3, 8, 4

    2022

  28. [36]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3

    2022

  29. [37]

    Introducing idefics: An open reproduction of state-of-the-art visual language model

    IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. https : / / huggingface.co/blog/idefics, 2023. 7

    2023

  30. [38]

    Oneformer: One transformer to rule universal image segmentation

    Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In CVPR, pages 2989– 2998, 2023. 6

    2023

  31. [39]

    Scaling up visual and vision-language representation learning with noisy text supervision

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904– 4916, 2021. 3

    2021

  32. [40]

    Vi- sual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Vi- sual prompt tuning. In ECCV, pages 709–727, 2022. 2, 3

    2022

  33. [41]

    Convolutional bypasses are better vision transformer adapters

    Shibo Jie and Zhi-Hong Deng. Convolutional bypasses are better vision transformer adapters. arXiv preprint arXiv:2207.07039, 2022. 3

    2022 arXiv

  34. [42]

    Fact: Factor-tuning for lightweight adaptation on vision transformer

    Shibo Jie and Zhi-Hong Deng. Fact: Factor-tuning for lightweight adaptation on vision transformer. InAAAI, pages 1060–1068, 2023. 2, 3

    2023

  35. [43]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. In ICCV, pages 4015–4026, 2023. 1, 3, 8, 4

    2023

  36. [44]

    Similarity of neural network representa- tions revisited

    Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representa- tions revisited. In ICML, pages 3519–3529, 2019. 3

    2019

  37. [45]

    Mask dino: Towards a unified transformer-based framework for object detection and segmentation

    Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In CVPR, pages 3041–3050, 2023. 6

    2023

  38. [46]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 7

    2023

  39. [47]

    Benchmarking detection transfer learning with vision transformers

    Yanghao Li, Saining Xie, Xinlei Chen, Piotr Dollar, Kaim- ing He, and Ross Girshick. Benchmarking detection transfer learning with vision transformers. arXiv preprint arXiv:2111.11429, 2021. 5, 7

    2021 arXiv

  40. [48]

    Exploring plain vision transformer backbones for object de- tection

    Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object de- tection. In ECCV, pages 280–296, 2022. 7

    2022

  41. [49]

    Evaluating object hallucina- tion in large vision-language models

    Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucina- tion in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 7

    2023 arXiv

  42. [50]

    Visual large language models for generalized and specialized applications

    Yifan Li, Zhixin Lai, Wentao Bao, Zhen Tan, Anh Dao, Kewei Sui, Jiayi Shen, Dong Liu, Huan Liu, and Yu Kong. Visual large language models for generalized and specialized applications. arXiv preprint arXiv:2501.02765, 2025. 7

    2025 arXiv

  43. [51]

    Binsformer: Revisiting adaptive bins for monocular depth estimation

    Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. IEEE TIP, 2024. 2, 9

    2024

  44. [52]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740–755, 2014. 5, 7

    2014

  45. [53]

    Va-depthnet: A variational approach to single image depth prediction

    Ce Liu, Suryansh Kumar, Shuhang Gu, Radu Timofte, and Luc Van Gool. Va-depthnet: A variational approach to single image depth prediction. In ICLR, 2023. 2

    2023

  46. [54]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296–26306, 2024. 7

    2024

  47. [55]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 7

    2024

  48. [56]

    Mmbench: Is your multi-modal model an all-around player? In ECCV, pages 216–233, 2024

    Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In ECCV, pages 216–233, 2024. 7

    2024

  49. [57]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012–10022, 2021. 5, 6, 7, 2

    2021

  50. [58]

    Swin transformer v2: Scaling up capacity and resolution

    Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, pages 12009–12019, 2022. 6, 2

    2022

  51. [59]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976–11986, 2022. 6, 7

    2022

  52. [60]

    Decoupled weight decay regularization

    I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5

    2017 arXiv

  53. [61]

    Learn to explain: Multimodal reasoning via thought chains for science question answering

    Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. NeurIPS,

  54. [62]

    Time-memory-and parameter-efficient visual adaptation

    Otniel-Bogdan Mercea, Alexey Gritsenko, Cordelia Schmid, and Anurag Arnab. Time-memory-and parameter-efficient visual adaptation. In CVPR, pages 5536–5545, 2024. 3, 5

    2024

  55. [63]

    The role of context for object detection and se- mantic segmentation in the wild

    Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and se- mantic segmentation in the wild. In CVPR, pages 891–898,

  56. [64]

    All in tokens: Uni- fying output space of visual tasks via soft token

    Jia Ning, Chen Li, Zheng Zhang, Chunyu Wang, Zigang Geng, Qi Dai, Kun He, and Han Hu. All in tokens: Uni- fying output space of visual tasks via soft token. In ICCV, pages 19900–19910, 2023. 2

    2023

  57. [65]

    Dinov2: Learning robust visual features without super- vision

    Maxime Oquab, Timoth ´ee Darcet, Th´eo Moutakanni, Huy V V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without super- vision. Transactions on Machine Learning Research,...

    2023

  58. [66]

    St-adapter: Parameter-efficient image-to-video transfer learning

    Junting Pan, Ziyi Lin, Xiatian Zhu, Jing Shao, and Hong- sheng Li. St-adapter: Parameter-efficient image-to-video transfer learning. In NeurIPS, pages 26462–26477, 2022. 3

    2022

  59. [67]

    P3depth: Monocular depth estimation with a piecewise planarity prior

    Vaishakh Patil, Christos Sakaridis, Alexander Liniger, and Luc Van Gool. P3depth: Monocular depth estimation with a piecewise planarity prior. In CVPR, pages 1610–1621, 2022. 2

    2022

  60. [68]

    idisc: Internal discretization for monocular depth estimation

    Luigi Piccinelli, Christos Sakaridis, and Fisher Yu. idisc: Internal discretization for monocular depth estimation. In CVPR, pages 21477–21487, 2023. 2

    2023

  61. [69]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In ICML, pages 8748–8763, 2021. 1, 3, 4

    2021

  62. [70]

    Do vision trans- formers see like convolutional neural networks? NeurIPS, 34:12116–12128, 2021

    Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision trans- formers see like convolutional neural networks? NeurIPS, 34:12116–12128, 2021. 3

    2021

  63. [71]

    Vi- sion transformers for dense prediction

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. In ICCV, pages 12179–12188, 2021. 2, 9

    2021

  64. [72]

    Learning multiple visual domains with residual adapters

    Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. In NeurIPS, 2017. 3

    2017

  65. [73]

    Iebins: Iterative elastic bins for monocular depth estimation

    Shuwei Shao, Zhongcai Pei, Xingming Wu, Zhong Liu, Wei- hai Chen, and Zhengguo Li. Iebins: Iterative elastic bins for monocular depth estimation. In NeurIPS, 2024. 2

    2024

  66. [74]

    Indoor segmentation and support inference from rgbd images

    Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, pages 746–760, 2012. 4

    2012

  67. [75]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105–6114, 2019. 2

    2019

  68. [76]

    Training data-efficient image transformers & distillation through at- tention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv ´e J´egou. Training data-efficient image transformers & distillation through at- tention. In ICML, pages 10347–10357, 2021. 3

    2021

  69. [77]

    Pvt v2: Improved baselines with pyramid vision transformer

    Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415–424, 2022. 5, 7

    2022

  70. [78]

    Internimage: Exploring large-scale vi- sion foundation models with deformable convolutions

    Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vi- sion foundation models with deformable convolutions. In CVPR, pages 14408–14419, 2023. 7

    2023

  71. [79]

    Vit-comer: Vision transformer with convolu- tional multi-scale feature interaction for dense predictions

    Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. Vit-comer: Vision transformer with convolu- tional multi-scale feature interaction for dense predictions. In CVPR, pages 5493–5502, 2024. 1, 2, 3, 5, 6, 7

    2024

  72. [80]

    Unified perceptual parsing for scene understand- ing

    Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understand- ing. In ECCV, pages 418–434, 2018. 2, 6

    2018

  73. [81]

    Focal self-attention for local-global interactions in vision transformers

    Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Xiyang Dai, Bin Xiao, Lu Yuan, and Jianfeng Gao. Focal self-attention for local-global interactions in vision transformers. arXiv preprint arXiv:2107.00641, 2021. 7

    2021 arXiv

  74. [82]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, pages 10371–10381, 2024. 1, 3

    2024

  75. [83]

    Visual tuning

    Bruce XB Yu, Jianlong Chang, Haixin Wang, Lingbo Liu, Shijie Wang, Zhiyu Wang, Junfan Lin, Lingxi Xie, Haojie Li, Zhouchen Lin, et al. Visual tuning. ACM Computing Surveys, 56(12):1–38, 2024. 3

    2024

  76. [84]

    Coca: Contrastive captioners are image-text foundation models

    Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mo- jtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 3

    2022

  77. [85]

    Mm-vet: Evaluating large multimodal models for integrated capabilities

    Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 7

    2023 arXiv

  78. [86]

    Neural window fully-connected crfs for monocular depth estimation

    Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocular depth estimation. In CVPR, pages 3916–3925, 2022. 2

    2022

  79. [87]

    Spanet: Frequency-balancing token mixer using spectral pooling aggregation modulation

    Guhnoo Yun, Juhan Yoo, Kijung Kim, Jeongho Lee, and Dong Hwan Kim. Spanet: Frequency-balancing token mixer using spectral pooling aggregation modulation. In ICCV, pages 6113–6124, 2023. 7

    2023

  80. [88]

    Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models

    Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In ACL, pages 1–9, 2022. 3

    2022

  81. [89]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975–11986, 2023. 3, 8, 4

    2023

  82. [90]

    Memory efficient transformer adapter for dense pre- dictions

    Dong Zhang, Rui Yan, Pingcheng Dong, and Kwang-Ting Cheng. Memory efficient transformer adapter for dense pre- dictions. In ICLR, 2025. 3

    2025

  83. [91]

    A survey of large language mod- els

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language mod- els. arXiv preprint arXiv:2303.18223, 2023. 3

    2023 arXiv

  84. [92]

    Semantic under- standing of scenes through the ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fi- dler, Adela Barriuso, and Antonio Torralba. Semantic under- standing of scenes through the ade20k dataset. IJCV, 127: 302–321, 2019. 5

    2019

  85. [93]

    A comprehensive survey on pretrained foundation mod- els: A history from bert to chatgpt

    Ce Zhou, Qian Li, Chen Li, Jun Yu, Yixin Liu, Guangjing Wang, Kai Zhang, Cheng Ji, Qiben Yan, Lifang He, et al. A comprehensive survey on pretrained foundation mod- els: A history from bert to chatgpt. arXiv preprint arXiv:2302.09419, 2023. 1

    2023 arXiv

  86. [94]

    Image bert pre-training with online tokenizer

    Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In ICLR, 2022. 1, 3

    2022

  87. [95]

    Conditional prompt learning for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zi- wei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816–16825, 2022. 3

    2022

  88. [96]

    Learning to prompt for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337–2348, 2022. 3

    2022

Pith tools

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