Pith. sign in

REVIEW 5 major objections 5 minor 45 references

MINR: Implicit Neural Representations with Masked Image Modelling

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper argues that replacing a masked autoencoder's discrete patch decoder with a continuous implicit neural representation, whose weights are predicted from the masked image, yields higher PSNR reconstruction at lower parameter…

desk verdict A clean TransINR/GINR-plus-MIM combination with strong reconstruction numbers, but the SSL utility claims rest entirely on the training loss itself. read the letter →

arxiv 2507.22404 v1 pith:BDYO2RSO submitted 2025-07-30 cs.CV cs.AIcs.LG

Sua Lee , Joonhun Lee , Myungjoo Kang This is my paper
classification cs.CVcs.AIcs.LG
keywords self-supervisedlearningmaskedimagemodelingimplicitneuralrepresentationsautoencodersout-of-distributiongeneralizationhypernetworkreconstructionPSNR
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

Masked autoencoders (MAE) reconstruct images by decoding visible patches into explicit predictions for masked ones, and their performance depends on what the visible patches supply. This paper proposes MINR, a framework that instead learns a continuous function mapping pixel coordinates to colors, with a transformer-based hypernetwork producing the function's weights from the masked image. On the three reported datasets, the method achieves higher reconstruction PSNR than MAE, about 6.4 dB higher on CelebA in-domain despite roughly 2.4 to 2.5 times fewer parameters, and keeps a lead of several dB when applying a model trained on one dataset to another. The intended payoff is a self-supervised pretraining framework that is less sensitive to masking strategy and to distribution shift, though the current evaluation stops at reconstruction quality.

What carries the argument

The central mechanism is the masked implicit neural representation (MINR): a transformer-based hypernetwork converts the visible patches of a masked image into the weights of a small coordinate MLP, which then evaluates the continuous function mapping pixel coordinates to RGB values at every location. The GINR variant splits the MLP into instance-agnostic layers plus one instance-specific layer, so the network learns dataset-wide structure and per-image detail separately. This replaces MAE's encoder-decoder over visible patches with a continuous function whose value at every masked coordinate comes from the full input condition rather than from explicit aggregation of visible-patch features; the paper argues this is why reconstruction quality no longer hinges on the amount of information the visible patches happen to contain.

What would settle it

Train MINR and MAE on identical masked-image data, then transfer each to a supervised classification task by linear probing or fine-tuning on a standard benchmark such as ImageNet or CIFAR-10. If MAE's downstream accuracy matches or exceeds MINR's while MINR has the higher reconstruction PSNR, the paper's central claim that the framework is a superior and robust alternative for self-supervised representation learning would be falsified.

Watch

Extended reading notes

Core claim

The central claim is that masked image modeling need not predict discrete patch contents from visible patches; the masked image can instead be treated as a conditioning signal that determines a continuous function $f_\theta: \mathbb{R}^2 \to \mathbb{R}^3$ mapping each pixel coordinate to an RGB value. MINR instantiates this with two INR backbones: TransINR, where the hypernetwork predicts the full weight set of a coordinate MLP, and GINR, where the second MLP layer is instance-specific while the rest are shared. In both settings, a 75 percent random mask is applied and the loss is computed over all pixels of the reconstructed coordinate function. The paper reports that this continuous reconstruction outperforms MAE in every in-domain and cross-domain comparison it runs, with about 44 million parameters versus MAE-Base's 106 million and MAE-Large's 314 million.

Load-bearing premise

The load-bearing premise is that the PSNR of masked-patch reconstruction measures the quality of the learned representation; if reconstruction fidelity does not imply good features for downstream tasks, the claimed advantage of MINR as a self-supervised learning framework would not be supported by the paper's experiments.

Editorial extensions

If this is right

  • If MINR's reconstruction advantage holds, masked-image pretraining can shift from discrete patch decoding to continuous coordinate functions, reducing dependence on heavy decoders and pretrained tokenizers.
  • Because the representation is a continuous function of pixel coordinates, a single trained model could in principle produce reconstructions or embeddings at arbitrary resolutions, an input-size flexibility the paper points to but does not yet test.
  • On out-of-distribution transfers, a model trained on one dataset and evaluated on another should retain higher reconstruction quality than MAE, which matters for adapting to unseen domains without retraining.
  • The parameter reduction (43.7 to 44.5 million versus 106.2 to 313.6 million for MAE) suggests cheaper pretraining and fine-tuning, assuming the reconstruction-quality advantage transfers to downstream tasks.

Reading between the lines

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

  • The paper does not run downstream linear-probe or fine-tuning experiments, so its strongest claim, that MINR is a robust self-supervised learning framework, remains inferred from reconstruction PSNR; comparing MINR and MAE on classification or detection after pretraining would test that inference directly.
  • Because the INR is resolution-agnostic, MINR could plausibly generalize to variable-size inputs or higher-resolution outputs without retraining, an extension the authors note but leave untested.
  • The claimed masking-strategy robustness could be stress-tested by evaluating MINR under mask ratios and mask shapes not seen in training, such as 90 percent masking or block-shaped masks; the paper motivates this claim but does not run that comparison.
  • If the GINR variant's instance-specific second layer is what carries cross-domain robustness, ablating which layer is instance-specific could clarify the mechanism and may yield further gains on other architectures.
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 / 5 minor

Summary. The manuscript proposes MINR, a masked image modeling framework that combines implicit neural representations (INRs) with masked-image pretraining. A transformer-based hypernetwork, instantiated as either TransINR or GINR, predicts INR weights from a masked image, and the INR maps pixel coordinates to RGB values, thereby reconstructing the image. The authors claim that the continuous functional form of INRs makes reconstruction less dependent on masking strategies and more robust to out-of-distribution data than MAE, while using fewer parameters. Experiments on CelebA, Imagenette, and MIT Indoor67 report PSNR for in-domain and cross-domain mask reconstruction, comparing MINR against MAE Base and Large, and the paper concludes that MINR is a robust and efficient self-supervised learning alternative.

Significance. If the representation-quality claim were established, the hybrid INR+MIM idea would be a useful contribution: the two backbone variants are clearly described, the parameter counts are substantially lower than the MAE baselines, and the cross-dataset transfer setup is a sensible first probe of robustness. However, as it stands the evidence supports only a claim about inpainting quality: all reported numbers are PSNR on the reconstruction task that MINR is directly trained for, and the advertised SSL utility is deferred to future work. The paper would be strengthened by downstream transfer experiments and masking-strategy sweeps; those are needed to substantiate the central claims. The manuscript does not provide code, but it states that official implementations and hyperparameters are used, which aids reproducibility of the reported baselines.

major comments (5)
  1. [§4.2-4.3, Tables 1-2] The central claim that MINR provides a superior self-supervised learning framework is not supported by the evaluation: the only metric is PSNR on mask reconstruction, which is essentially the L2 objective minimized in Eq. (2) for MINR, while MAE is trained only on masked patches and is not designed to maximize full-image reconstruction PSNR. Since the abstract and introduction promise 'robust feature representations' and 'flexibility in creating embeddings for downstream tasks,' the paper needs at least one downstream evaluation (e.g., linear probing or fine-tuning for classification or semantic segmentation) to connect reconstruction quality to representation quality; the concluding sentence that 'In the future, we plan to leverage the flexibility of MINR ... to showcase its performance ... across various downstream tasks' confirms that this evaluation is currently absent.
  2. [Eq. (2), §4.3] The PSNR computation is ambiguous and potentially biased. Eq. (2) sums over all HxW pixels, whereas MAE's loss is computed on masked patches only; if the Tables report full-image PSNR, MINR can obtain a trivial advantage by reconstructing the visible patches it was trained on, while MAE never optimizes those pixels. Please state explicitly whether the reported PSNR is computed on the masked region only, on the full image, or on the masked region after pasting visible patches, and report masked-region PSNR in addition to any full-image number.
  3. [§4.2] The claim of robustness 'irrespective of masking strategies' is not tested: all experiments use a single masking setting (75% random masks of 14x14 patches). The paper should include a sweep over mask ratios (e.g., 50%, 75%, 90%), mask sizes, and masking patterns (random, block, low-level) to substantiate the claim; otherwise the stated advantage over MAE's sensitivity to masking is an untested assertion.
  4. [Tables 1-2] The comparison to MAE is not fully specified: it is unclear whether MAE is pretrained on ImageNet and then fine-tuned on each dataset, trained from scratch on each dataset, or used with official checkpoints, and whether MINR is always trained from scratch on the source dataset only. Without identical training budgets, data splits, and initialization protocols, the PSNR differences and parameter-count comparison in Tables 1-2 cannot be interpreted as intrinsic advantages. Please specify the training protocol and report standard deviations over seeds, since every reported number is currently a single PSNR value without error bars.
  5. [Table 2] In the OOD setting, the paper reports transfer between datasets but does not describe whether the target images are seen only at test time or whether any fine-tuning occurs; additionally, PSNR improvement on natural images from a different distribution does not by itself demonstrate that the learned features are more robust for downstream tasks. An OOD linear-probe or fine-tuning evaluation would directly test the representation robustness claimed in the introduction.
minor comments (5)
  1. [§4.1, references [20] and [29]] The references for CelebA and Imagenette appear to be swapped: CelebA is the dataset of Liu et al. (2015), not the fastai book, and Imagenette is the fastai dataset, not the face-attributes paper.
  2. [Abstract] The phrase 'image reconstruction-based pretraining task' should be made grammatical, e.g., 'image reconstruction-based pretraining tasks'.
  3. [§4.2] The input resolution of 182x182 differs from the standard MAE input resolution; please justify this choice and comment on its effect on the comparison.
  4. [Figure 1] The figure does not state which dataset or masking configuration is shown, and the images are too small to judge reconstruction quality; adding labels and enlarged crops would help.
  5. [§5] The sentence 'demonstrated the superiority ... as substantiated by higher PSNR values across different datasets' overstates the evidence, since only one metric and one masking configuration are reported.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the PSNR/objective overlap is a construct-validity gap, not a circular reduction.

full rationale

The core empirical claim—MINR outperforms MAE on in-domain and cross-domain mask reconstruction with fewer parameters—is tested against an external baseline (MAE) on held-out data from CelebA, Imagenette, and MIT Indoor67 (Tables 1 and 2). These comparisons are falsifiable and are not determined by MINR's training objective; the model could in principle lose to MAE. The fact that PSNR is a monotone transform of the L2 loss minimized in Eq. (2) means the paper's metric overlaps with its training objective, but that overlap would be circular only if the paper derived its MAE-beating result from the loss itself; it does not. The broader SSL claim in the abstract and conclusion ('versatility ... various self-supervised learning applications', 'robust feature representations') is not supported by any downstream linear-probe or fine-tuning experiment, and the 'robustness to diverse masking strategies' claim is tested under only one masking configuration (75% random 14×14 patches, §4.2). These are evidence/construct-validity gaps, not circular reductions. The architecture backbones are external prior works ([12, 23]), and no load-bearing self-citation chain appears; the conclusion explicitly defers downstream-task evaluation to future work.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on external architectural components (TransINR, GINR, ViT-style hypernetwork) and on empirical assumptions about PSNR as an SSL proxy. There are no new free parameters or invented entities, but several fixed hyperparameters are not varied despite a robustness claim.

free parameters (4)
  • mask_ratio = 0.75
    Chosen following MAE convention and not varied in experiments, despite the paper's claim of robustness to masking strategies.
  • patch_size = 14x14
    Chosen to match the 182x182 input resolution; no ablation is performed.
  • MLP_depth = 5 layers
    Architecture choice for the coordinate-based MLP; not justified experimentally.
  • instance_specific_layer_index = 2
    Taken from GINR's empirically most performant configuration; not re-validated in this paper.
assumptions (4)
  • domain assumption MAE's reconstruction quality depends on the quantity and quality of visible patches and degrades with unseen masking strategies.
    Cited from prior work [7,24] and used to motivate MINR, but not directly tested in this paper.
  • ad hoc to paper INRs are inherently less affected by variations in visible-patch information than MAE's discrete token reconstruction.
    Central motivation asserted in Sections 1 and 3.1; no derivation or controlled experiment isolates this effect.
  • domain assumption Mask-reconstruction PSNR is a valid proxy for self-supervised representation quality.
    Used as the only evaluation metric; no downstream linear-probe or fine-tuning results are provided.
  • domain assumption Cross-dataset PSNR transfer measures out-of-distribution generalization for self-supervised learning.
    The paper equates OOD robustness with reconstruction PSNR on a different dataset, without calibrating to actual downstream OOD tasks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MINR: Implicit Neural Representations with Masked Image Modelling." pith.science (2026). https://pith.science/paper/BDYO2RSO

@misc{pith2026250722404,
  author       = {Pith},
  title        = {Pith review of: MINR: Implicit Neural Representations with Masked Image Modelling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BDYO2RSO}},
  note         = {Machine review of arXiv:2507.22404}
}
read the original abstract

Self-supervised learning methods like masked autoencoders (MAE) have shown significant promise in learning robust feature representations, particularly in image reconstruction-based pretraining task. However, their performance is often strongly dependent on the masking strategies used during training and can degrade when applied to out-of-distribution data. To address these limitations, we introduce the masked implicit neural representations (MINR) framework that synergizes implicit neural representations with masked image modeling. MINR learns a continuous function to represent images, enabling more robust and generalizable reconstructions irrespective of masking strategies. Our experiments demonstrate that MINR not only outperforms MAE in in-domain scenarios but also in out-of-distribution settings, while reducing model complexity. The versatility of MINR extends to various self-supervised learning applications, confirming its utility as a robust and efficient alternative to existing frameworks.

Figures

Figures reproduced from arXiv: 2507.22404 by the authors.

Figure 1
Figure 1. Qualitative results of mask reconstruction. For each row, we present the masked image, MAE and MINR reconstructions, and the ground truth, in sequence. main, including domain generalization scenarios [1, 19]. Notable SSL techniques include masked image modelling (MIM) that enhance representation robustness by deliber￾ately masking images and training models to reconstruct the hidden information. The effectiveness of… view at source ↗
Figure 2
Figure 2. A schematic illustration of MINR. During the training phase, a large random subset of image patches are masked out. To ensure robustness, we employ a transformer-based hypernetwork predicting weights for an INR. TransINR directly maps weight sets [12], whereas GINR modulates only the second MLP layer as instance-specific, keeping the rest instance-agnostic [23]. The overall framework constructs weights with the mask… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 35 canonical work pages

  1. [1]

    Improving out-of-distribution gener- alization via multi-task self-supervised pretraining

    Isabela Albuquerque, Nikhil Naik, Junnan Li, Nitish Keskar, and Richard Socher. Improving out-of-distribution gener- alization via multi-task self-supervised pretraining. arXiv preprint arXiv:2003.13525, 2020. 1

    arXiv 2003

  2. [2]

    Image generators with conditionally-independent pixel synthesis

    Ivan Anokhin, Kirill Demochkin, Taras Khakhulin, Gleb Sterkin, Victor Lempitsky, and Denis Korzhenkov. Image generators with conditionally-independent pixel synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14278–14287, 2021. 3

    work page 2021

  3. [3]

    Efficient self-supervised learning with contextualized target representations for vision, speech and language

    Alexei Baevski, Arun Babu, Wei-Ning Hsu, and Michael Auli. Efficient self-supervised learning with contextualized target representations for vision, speech and language. In In- ternational Conference on Machine Learning , pages 1416–

    work page

  4. [4]

    Data2vec: A general frame- work for self-supervised learning in speech, vision and lan- guage

    Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general frame- work for self-supervised learning in speech, vision and lan- guage. In International Conference on Machine Learning , pages 1298–1312. PMLR, 2022. 2

    work page 2022

  5. [5]

    Masked autoencoders enable efficient knowledge distillers

    Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie. Masked autoencoders enable efficient knowledge distillers. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24256–24265, 2023. 2

    work page 2023

  6. [6]

    Beit: Bert pre-training of image transformers

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2

    arXiv 2021

  7. [7]

    How to understand masked autoencoders

    Shuhao Cao, Peng Xu, and David A Clifton. How to understand masked autoencoders. arXiv preprint arXiv:2202.03670, 2022. 2

    arXiv 2022

  8. [8]

    Pimae: Point cloud and image interactive masked autoencoders for 3d object detection

    Anthony Chen, Kevin Zhang, Renrui Zhang, Zihan Wang, Yuheng Lu, Yandong Guo, and Shanghang Zhang. Pimae: Point cloud and image interactive masked autoencoders for 3d object detection. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 5291–5301, 2023. 1

    work page 2023

Show all 45 references
  1. [9]

    Improving masked autoencoders by learning where to mask

    Haijian Chen, Wendong Zhang, Yunbo Wang, and Xiaokang Yang. Improving masked autoencoders by learning where to mask. arXiv preprint arXiv:2303.06583, 2023. 2

    2023 arXiv

  2. [10]

    Learning continuous image representation with local implicit image function

    Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628–8638,

  3. [11]

    Sdae: Self- distillated masked autoencoder

    Yabo Chen, Yuchen Liu, Dongsheng Jiang, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. Sdae: Self- distillated masked autoencoder. In European Conference on Computer Vision, pages 108–124. Springer, 2022. 2

    2022

  4. [12]

    Transformers as meta- learners for implicit neural representations

    Yinbo Chen and Xiaolong Wang. Transformers as meta- learners for implicit neural representations. In European Conference on Computer Vision , pages 170–187. Springer,

  5. [13]

    Image super-resolution using deep convolutional net- works

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Image super-resolution using deep convolutional net- works. IEEE transactions on pattern analysis and machine intelligence, 38(2):295–307, 2015. 4

    2015

  6. [14]

    Bootstrapped masked autoencoders for vision bert pretraining

    Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, and Nenghai Yu. Bootstrapped masked autoencoders for vision bert pretraining. In European Conference on Computer Vi- sion, pages 247–264. Springer, 2022. 2

    2022

  7. [15]

    Maskclip: Masked self- distillation advances contrastive language-image pretraining

    Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self- distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision ...

    2023

  8. [16]

    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. arXiv preprint a...

    2010 arXiv

  9. [17]

    From data to functa: Your data point is a function and you can treat it like one

    Emilien Dupont, Hyunjik Kim, SM Eslami, Danilo Rezende, and Dan Rosenbaum. From data to functa: Your data point is a function and you can treat it like one. arXiv preprint arXiv:2201.12204, 2022. 3

    2022 arXiv

  10. [18]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 16000– 16009, 2022. 2

    2022

  11. [19]

    Using self-supervised learning can improve model robustness and uncertainty

    Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. Advances in neural in- formation processing systems, 32, 2019. 1

    2019

  12. [20]

    Fastai: A layered api for deep learning

    Jeremy Howard and Sylvain Gugger. Fastai: A layered api for deep learning. Information, 11(2):108, 2020. 3

    2020

  13. [21]

    Green hierarchical vision transformer for masked image modeling.Advances in Neural Information Processing Systems, 35:19997–20010, 2022

    Lang Huang, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, and Toshihiko Yamasaki. Green hierarchical vision transformer for masked image modeling.Advances in Neural Information Processing Systems, 35:19997–20010, 2022. 2

    2022

  14. [22]

    What to hide from your students: Attention-guided masked image modeling

    Ioannis Kakogeorgiou, Spyros Gidaris, Bill Psomas, Yan- nis Avrithis, Andrei Bursuc, Konstantinos Karantzalos, and Nikos Komodakis. What to hide from your students: Attention-guided masked image modeling. In European Conference on Computer Vision , pages 300–318. Springer,

  15. [23]

    Generalizable implicit neural represen- tations via instance pattern composers

    Chiheon Kim, Doyup Lee, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Generalizable implicit neural represen- tations via instance pattern composers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11808–11817, 2023. 2, 3

    2023

  16. [24]

    Un- derstanding masked autoencoders via hierarchical latent variable models

    Lingjing Kong, Martin Q Ma, Guangyi Chen, Eric P Xing, Yuejie Chi, Louis-Philippe Morency, and Kun Zhang. Un- derstanding masked autoencoders via hierarchical latent variable models. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages...

    2023

  17. [25]

    Exploring the role of mean teach- ers in self-supervised masked auto-encoders

    Youngwan Lee, Jeffrey Willette, Jonghee Kim, Juho Lee, and Sung Ju Hwang. Exploring the role of mean teach- ers in self-supervised masked auto-encoders. arXiv preprint arXiv:2210.02077, 2022. 2

    2022 arXiv

  18. [26]

    Semmae: Semantic-guided mask- ing for learning masked autoencoders

    Gang Li, Heliang Zheng, Daqing Liu, Chaoyue Wang, Bing Su, and Changwen Zheng. Semmae: Semantic-guided mask- ing for learning masked autoencoders. Advances in Neural Information Processing Systems, 35:14290–14302, 2022. 2

    2022

  19. [27]

    Architecture-agnostic masked image modeling–from vit back to cnn

    Siyuan Li, Di Wu, Fang Wu, Zelin Zang, Baigui Sun, Hao Li, Xuansong Xie, Stan Li, et al. Architecture-agnostic masked image modeling–from vit back to cnn. arXiv preprint arXiv:2205.13943, 2022. 2

    2022 arXiv

  20. [28]

    Mst: Masked self-supervised transformer for visual representation

    Zhaowen Li, Zhiyang Chen, Fan Yang, Wei Li, Yousong Zhu, Chaoyang Zhao, Rui Deng, Liwei Wu, Rui Zhao, Ming Tang, et al. Mst: Masked self-supervised transformer for visual representation. Advances in Neural Information Pro- cessing Systems, 34:13165–13176, 2021. 2

    2021

  21. [29]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730–3738, 2015. 3

    2015

  22. [30]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM , 65(1):99–106, 2021. 3

    2021

  23. [31]

    Towards un- derstanding why mask reconstruction pretraining helps in downstream tasks

    Jiachun Pan, Pan Zhou, and Y AN Shuicheng. Towards un- derstanding why mask reconstruction pretraining helps in downstream tasks. In The Eleventh International Conference on Learning Representations, 2022. 2

    2022

  24. [32]

    Beit v2: Masked image modeling with vector-quantized visual tokenizers

    Zhiliang Peng, Li Dong, Hangbo Bao, Qixiang Ye, and Furu Wei. Beit v2: Masked image modeling with vector-quantized visual tokenizers. arXiv preprint arXiv:2208.06366, 2022. 2

    2022 arXiv

  25. [33]

    Recognizing indoor scenes

    Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In 2009 IEEE conference on computer vision and pattern recognition, pages 413–420. IEEE, 2009. 3

    2009

  26. [34]

    Adversarial masking for self-supervised learning

    Yuge Shi, N Siddharth, Philip Torr, and Adam R Kosiorek. Adversarial masking for self-supervised learning. In Inter- national Conference on Machine Learning , pages 20026– 20040. PMLR, 2022. 2

    2022

  27. [35]

    Learned initializations for optimizing coordinate-based neural representations

    Matthew Tancik, Ben Mildenhall, Terrance Wang, Divi Schmidt, Pratul P Srinivasan, Jonathan T Barron, and Ren Ng. Learned initializations for optimizing coordinate-based neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2021

  28. [36]

    Hard patches mining for masked image modeling

    Haochen Wang, Kaiyou Song, Junsong Fan, Yuxi Wang, Jin Xie, and Zhaoxiang Zhang. Hard patches mining for masked image modeling. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 10375–10385, 2023. 2

    2023

  29. [37]

    Masked feature predic- tion for self-supervised visual pre-training

    Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature predic- tion for self-supervised visual pre-training. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14668–14678, 2022. 2

    2022

  30. [38]

    Con- vnext v2: Co-designing and scaling convnets with masked autoencoders

    Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Con- vnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133– 161...

    2023

  31. [39]

    Dropmae: Masked autoen- coders with spatial-attention dropout for tracking tasks

    Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B Chan. Dropmae: Masked autoen- coders with spatial-attention dropout for tracking tasks. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 14561–14571, 2023. 1

    2023

  32. [40]

    Denoising masked autoen- coders are certifiable robust vision learners

    Quanlin Wu, Hang Ye, Yuntian Gu, Huishuai Zhang, Liwei Wang, and Di He. Denoising masked autoen- coders are certifiable robust vision learners. arXiv preprint arXiv:2210.06983, 2022. 2

    2022 arXiv

  33. [41]

    Ex- treme masking for learning instance and distributed visual representations

    Zhirong Wu, Zihang Lai, Xiao Sun, and Stephen Lin. Ex- treme masking for learning instance and distributed visual representations. arXiv preprint arXiv:2206.04667, 2022. 2

    2022 arXiv

  34. [42]

    Masked frequency model- ing for self-supervised visual pre-training

    Jiahao Xie, Wei Li, Xiaohang Zhan, Ziwei Liu, Yew Soon Ong, and Chen Change Loy. Masked frequency model- ing for self-supervised visual pre-training. arXiv preprint arXiv:2206.07706, 2022. 2

    2022 arXiv

  35. [43]

    Simmim: A simple framework for masked image modeling

    Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9653–9663, 2022. 2

    2022

  36. [44]

    Moma: Distill from self-supervised teachers

    Yuchong Yao, Nandakishor Desai, and Marimuthu Palaniswami. Moma: Distill from self-supervised teachers. arXiv preprint arXiv:2302.02089, 2023. 2

    2023 arXiv

  37. [45]

    Mimco: Masked image modeling pre-training with con- trastive teacher

    Qiang Zhou, Chaohui Yu, Hao Luo, Zhibin Wang, and Hao Li. Mimco: Masked image modeling pre-training with con- trastive teacher. In Proceedings of the 30th ACM Interna- tional Conference on Multimedia , pages 4487–4495, 2022. 2

    2022

Pith tools

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