Pith. sign in

REVIEW 5 major objections 5 minor 40 references

Perception-Oriented Latent Coding for High-Performance Compressed Domain Semantic Inference

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 Fine-tuning only a small adapter, a perception-trained codec outperforms fully fine-tuned MSE codecs on vision tasks.

desk verdict A competent empirical systems paper whose central transferability claim is narrower than advertised; solid within its tested ImageNet/ADE20K scope, but the semantic-richness story needs a task far from ImageNet to convince. read the letter →

arxiv 2507.01608 v1 pith:CJ3BFQYV submitted 2025-07-02 cs.CV eess.IV

Xu Zhang , Ming Lu , Yan Chen , Zhan Ma This is my paper · ORCID
classification cs.CVeess.IV
keywords learnedimagecodingcompresseddomainsemanticinferenceperception-orientedoptimizationgenerativeadversarialnetworksLPIPSadapterfine-tuninglatentspacediscriminabilityNetclassification
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

This paper tries to establish that the training objective of a learned image coder determines whether its compressed latent code can be reused directly for downstream vision tasks. It introduces Perception-Oriented Latent Coding (POLC), which trains the coder with adversarial and deep-feature perceptual losses in addition to rate and MSE terms, and claims the resulting latent space is semantically rich enough that only a small plug-and-play adapter needs fine-tuning for ImageNet classification and ADE20K segmentation. The paper reports that adapter-only fine-tuning of POLC outperforms fully fine-tuned models with MSE-optimized codecs while keeping reconstruction quality comparable to state-of-the-art generative image coders. A sympathetic reader would care because compressed-domain inference promises to skip full image decoding, and reducing fine-tuning cost makes this practical for large vision models.

What carries the argument

The carrying mechanism is the perception-oriented training objective of Equations (1)-(4): a rate loss, a reconstruction loss combining MSE with LPIPS, and a GAN generator loss under a conditional discriminator. LPIPS, computed from deep network features, aligns the latent representation with semantic-level similarity, while the adversarial loss encourages the decoder to produce perceptually faithful images; together they make the compressed latent $\hat{z}$ more discriminative. The adapter is then a small module made of a linear channel projection, a pixel-shuffle spatial alignment, and a residual block, trained with a task loss plus a feature-distillation loss from the frozen teacher vision model.

What would settle it

Train POLC and an MSE-optimized codec on the same data, then evaluate adapter-only fine-tuning on a downstream task whose categories are far from ImageNet, such as cell counting, defect detection, or satellite scene classification; if adapter-only POLC no longer matches or beats fully fine-tuned MSE baselines, the central transferability claim fails.

Watch

Extended reading notes

Core claim

The central claim is that replacing the pure MSE objective of learned image coding with a perception-oriented objective changes the geometry of the compressed latent space: samples of the same class become closer together and different classes become more separated, so an off-the-shelf vision backbone can be adapted with only a lightweight channel-and-resolution adapter instead of full fine-tuning. With the image coder frozen, the authors fine-tune only this universal adapter on ImageNet-1K classification and ADE20K semantic segmentation, and report that it outperforms fully fine-tuned models built on MSE-optimized codecs, reaches accuracy close to the pixel-domain method MPA, and achieves rate-perception performance comparable to HiFiC and MS-ILLM.

Load-bearing premise

The paper assumes that LPIPS and the GAN objective make the compressed latent semantically discriminative for arbitrary downstream tasks, but the only evidence covers ImageNet classification and ADE20K segmentation, both close to the supervision used to train the perceptual and distillation features.

Editorial extensions

If this is right

  • Adapter-only fine-tuning may suffice for global understanding tasks such as classification, while dense prediction tasks like segmentation can still benefit from fully fine-tuning the vision model on top of POLC.
  • Larger vision backbones appear to gain more from POLC, as accuracy improves when moving from ConvNeXt-T to ConvNeXt-L.
  • Compared with MSE-oriented compressed-domain models, POLC cuts fine-tuning parameters from about 29.19M to 0.60M for ConvNeXt-T while improving Top-1 accuracy at a comparable bitrate.
  • Bit allocation shifts away from high-frequency texture peaks toward semantic regions, which is consistent with the claim that the latent code carries object-level information rather than just pixel fidelity.

Reading between the lines

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

  • The transferability claim is only demonstrated for tasks aligned with ImageNet semantics, so the same adapter-only advantage may not appear for detection, counting, or medical and remote sensing tasks whose semantic axes are absent from the LPIPS teacher features.
  • The UMAP-based claim of a more discriminative latent space could be quantified directly, for example by linear-probe accuracy or k-nearest-neighbor purity of the compressed latents, which would test whether semantic richness precedes fine-tuning rather than being created by it.
  • If the perceptual objective is what creates the advantage, then a similar training shift could be applied to learned video or point-cloud codecs that also serve downstream machine perception, though the paper does not test those settings.
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 paper proposes POLC, a perception-oriented training scheme for learned image coders that combines GAN-based generative coding, an LPIPS perceptual loss, and an adapter-based fine-tuning procedure with feature distillation, with the goal of producing compressed latents that are directly usable for downstream semantic inference. The main claims are (i) that POLC achieves reconstruction rate-perception quality comparable to HiFiC and MS-ILLM, and (ii) that fine-tuning only a small universal adapter on POLC latents outperforms full fine-tuning of MSE-optimized LIC models on ImageNet-1K classification and ADE20K segmentation while using far fewer trainable parameters. The paper also includes UMAP visualizations, bit-allocation analyses, and a complexity comparison to support the proposed mechanism.

Significance. If the claims hold, the paper offers a practical recipe for compressed-domain semantic inference that separates a frozen compressor from an inexpensive task adapter, potentially reducing the adaptation cost for large vision models in machine-vision applications. The evaluation is broad: three vision backbones (ConvNeXt, DeiT, PSPNet) across two tasks, rate-distortion and rate-perception plots, complexity/latency analysis, and ablations of adapter components. However, the core 'semantic richness' mechanism is only tested on ImageNet-aligned tasks, and several quantitative supports are missing, including error bars, a complete complexity table, and numeric rate-perception comparisons. The contribution is incremental but potentially useful to the compressed-domain inference community; the claimed universality currently exceeds the provided evidence.

major comments (5)
  1. [§IV-B, Fig. 4] All reported task accuracy curves are single-run values with no error bars, seed variance, or statistical significance tests. The central claim that 'fine-tuning only the adapter, POLC outperforms fully fine-tuned models with MSE-optimized LIC' rests on differences of roughly 3–4 percentage points at several bitrates; without variance estimates, the reader cannot determine whether these gaps exceed run-to-run noise. Please report mean ± std over at least three seeds for the main comparisons, or at least at a representative bitrate such as 0.1 bpp.
  2. [§IV-C, Table I] Table I is malformed: the POLC row lists only two numeric entries ('0.60M (-28.59M) 76.54 (+3.48)') against four column headers (#Params for Ft., GFLOPs, Latency (ms), Acc. @ 0.1bpp), so neither the latency nor the accuracy of POLC at 0.1 bpp is reported. This table is the primary quantitative evidence for the claim of 'minimal fine-tuning overhead with comparable accuracy,' and the omission must be fixed. Please also state whether the 'MSE' row corresponds to full fine-tuning and whether the numbers are measured or taken from prior work.
  3. [§III-A, Eq. (3); §V] The semantic-richness claim is only tested on ImageNet-1K classification and ADE20K segmentation, both closely aligned with the supervision cues used in training: LPIPS in Eq. (3) uses features from an ImageNet-trained network, and the distillation loss in Eq. (6) uses the same vision backbones that are later evaluated on the same tasks. This makes the 'perception-oriented latent is semantically rich for arbitrary downstream tasks' framing broader than the evidence. Please add at least one experiment on a task whose semantics are not part of ImageNet category structure (e.g., monocular depth estimation, surface-normal estimation, or instance retrieval) under the same adapter protocol, or explicitly scope the claim to tasks whose semantics are representable by ImageNet/LPIPS features. Without this, the mechanism's generality is unsubstantiated.
  4. [§IV-B, Fig. 4] The claim that POLC achieves 'rate-perception performance comparable to state-of-the-art generative image coding methods' is supported only by qualitative curve overlap with HiFiC and MS-ILLM; no numerical rate-perception metrics (e.g., BD-rate computed with LPIPS or FID, or FID/LPIPS values at matched bitrates) or error bars are provided. Since this is one of the two headline contributions, please add numeric comparisons to the R-P curves.
  5. [§IV-C, Fig. 1] The UMAP visualization of 10 ImageNet classes is qualitative; no quantitative metric (e.g., k-NN accuracy, silhouette score, or mutual information) is provided to substantiate the claim that POLC produces a 'more discriminative latent space.' Because this is the core mechanism motivating the entire approach, please add a quantitative measure of latent discriminability for both POLC and MSE.
minor comments (5)
  1. [Eq. (4)] The symbol LEGP is introduced but never defined; please spell out the acronym (e.g., 'encoder-generator-perception loss') and state explicitly that it is the total loss used to train the coder.
  2. [Eq. (3)] Lperc is mentioned in the text and used in Eq. (3), but the definition is only implicit. Please write Lperc = LPIPS(x, x̂) explicitly to avoid ambiguity.
  3. [§IV-A] The training paragraph lists λtask = 1 among loss coefficients, but Eq. (4) does not include a task loss; if λtask is only used in adapter training, remove it from the coder coefficient list to avoid confusion.
  4. [Table I] The ablation rows 'Shuffle→TConv' and 'w/o ResBlock' are reported as deltas; please specify the baseline to which these deltas apply (presumably the full POLC row) and ensure all rows have the same number of entries.
  5. [Conclusion] There is a typo: 'differrent' should be 'different,' and 'design choices is' should be 'design choices are.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical comparisons against external baselines, and the LPIPS/ImageNet alignment is a generalization limitation rather than a circular reduction.

full rationale

The paper's derivation chain is empirical rather than formal: POLC is trained with Eqs. (1)-(4) using a GAN objective, LPIPS perceptual loss, rate loss, and MSE reconstruction, and the downstream adapter is trained with Eqs. (5)-(7) using task loss and feature distillation. The headline claim, that adapter-only fine-tuning of POLC outperforms fully fine-tuned MSE-oriented LIC, is tested on held-out ImageNet-1K and ADE20K splits against external baselines such as HiFiC, MS-ILLM, MPA, Adapt-ICMH, TransTIC, and VTM. No equation is defined in terms of an evaluation quantity, and no fitted parameter is renamed as a prediction. The use of LPIPS in Eq. (3), whose deep features come from an ImageNet-trained network, means that ImageNet and ADE20K results are partly aligned with the training objective; however, this is a generalization limitation rather than a circular reduction, because LPIPS distance and downstream task accuracy are distinct quantities and the comparison against MSE-trained latents is an empirical result. The self-citations, including MPA [16] and UAC [21], provide training recipes, distillation baselines, and design choices, but the central comparisons are against externally published methods and standard benchmarks, so these self-citations are not load-bearing. The stated limitation in the Conclusion, that the adapter must be retrained for each vision model, affects generality but does not indicate circularity. No circular step was found.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities. Its claims rest on tuned loss coefficients and on domain assumptions about LPIPS, GAN training, frozen-backbone transfer, and distillation weighting. The most consequential assumption is that ImageNet-derived perceptual features generalize as a source of semantic richness.

free parameters (6)
  • generator loss weight lambda_G = 0.8
    Generator loss weight in Eq. 4, chosen by hand; no sensitivity analysis is reported.
  • distortion loss weight lambda_d = 1
    Distortion weight in Eq. 3, fixed to 1.
  • perceptual loss weight lambda_perc = 1
    LPIPS weight in Eq. 3, fixed to 1.
  • task loss weight lambda_task = 1
    Task loss weight in Eq. 7, fixed.
  • distillation loss weight lambda_dist = 0.001 (ConvNeXt-L), 0.01 (ConvNeXt-T), 0.1 (DeiT-S), 10 (PSPNet50)
    Tuned per model and architecture; the wide range suggests sensitivity.
  • rate weight lambda_rate = sampled from {18.0, 9.32, 4.83, 2.5, 1.3, 0.67, 0.35, 0.18}
    Controls the rate-distortion tradeoff; follows the variable-rate setting of MPA.
assumptions (4)
  • domain assumption Lperc = LPIPS(x, xhat) is a valid proxy for semantic-level perception and improves latent discriminability.
    Eq. 3 in Section III-A; LPIPS is a deep-feature distance from an ImageNet-trained network, so using it injects task-relevant priors into the codec.
  • domain assumption GAN training with discriminator D and generator loss LG enriches latent semantics, not just reconstruction realism.
    Eqs. 1, 2, and 4; adopted from HiFiC; no analysis isolates the GAN contribution to downstream task accuracy.
  • domain assumption With LIC and vision backbone frozen, training only the adapter with Ltask plus Ldist preserves sufficient task information.
    Section III-B training strategy; relies on the latent code already containing transferable semantics.
  • domain assumption A single shared lambda_i = 1 for all feature distillation layers is adequate.
    Section III-B after Eq. 6; inherited from reference [21], no sensitivity study is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Perception-Oriented Latent Coding for High-Performance Compressed Domain Semantic Inference." pith.science (2026). https://pith.science/paper/CJ3BFQYV

@misc{pith2026250701608,
  author       = {Pith},
  title        = {Pith review of: Perception-Oriented Latent Coding for High-Performance Compressed Domain Semantic Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CJ3BFQYV}},
  note         = {Machine review of arXiv:2507.01608}
}
read the original abstract

In recent years, compressed domain semantic inference has primarily relied on learned image coding models optimized for mean squared error (MSE). However, MSE-oriented optimization tends to yield latent spaces with limited semantic richness, which hinders effective semantic inference in downstream tasks. Moreover, achieving high performance with these models often requires fine-tuning the entire vision model, which is computationally intensive, especially for large models. To address these problems, we introduce Perception-Oriented Latent Coding (POLC), an approach that enriches the semantic content of latent features for high-performance compressed domain semantic inference. With the semantically rich latent space, POLC requires only a plug-and-play adapter for fine-tuning, significantly reducing the parameter count compared to previous MSE-oriented methods. Experimental results demonstrate that POLC achieves rate-perception performance comparable to state-of-the-art generative image coding methods while markedly enhancing performance in vision tasks, with minimal fine-tuning overhead. Code is available at https://github.com/NJUVISION/POLC.

Figures

Figures reproduced from arXiv: 2507.01608 by the authors.

Figure 1
Figure 1. Latent space visualization of MSE- and perception-oriented optimiza [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Perception-oriented latent coding. It leverages the generative image [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Universal adapter. With POLC, only the adapter need to be trained [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Reconstruction quality on the Kodak [33] and CLIC test set [34] and vision task performance on ImageNet-1K [18] and ADE20K [35]. While achieving [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Visualization of reconstruction and bit allocation. Compared to MSE [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 37 canonical work pages

  1. [1]

    The JPEG still picture compression standard,

    Gregory K Wallace, “The JPEG still picture compression standard,” Communications of the ACM , vol. 34, no. 4, pp. 30–44, 1991

    work page 1991

  2. [2]

    Bpg image format,

    Fabrice Bellard, “Bpg image format,” 2014

    work page 2014

  3. [3]

    Versatile video coding,

    ITU-T and ISO/IEC, “Versatile video coding,” ITU-T Rec. H.266 and ISO/IEC 23090-3 , 2020

    work page 2020

  4. [4]

    End-to-end optimized image compression,

    Johannes Ball ´e, Valero Laparra, and Eero P. Simoncelli, “End-to-end optimized image compression,” in International Conference on Learning Representations, 2017

    work page 2017

  5. [5]

    Variational image compression with a scale hyperprior,

    Johannes Ball ´e, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston, “Variational image compression with a scale hyperprior,” in International Conference on Learning Representations , 2018

    work page 2018

  6. [6]

    Joint autore- gressive and hierarchical priors for learned image compression,

    David Minnen, Johannes Ball ´e, and George D Toderici, “Joint autore- gressive and hierarchical priors for learned image compression,” in Advances in Neural Information Processing Systems . 2018, vol. 31, pp. 10794–10803, Curran Associates, Inc

    work page 2018

  7. [7]

    Transformer-based image compression,

    Ming Lu, Peiyao Guo, Huiqing Shi, Chuntong Cao, and Zhan Ma, “Transformer-based image compression,” in 2022 Data Compression Conference (DCC), 2022, pp. 469–469

    work page 2022

  8. [8]

    High-Efficiency Lossy Image Coding Through Adaptive Neighborhood Information Aggregation

    Ming Lu, Fangdong Chen, Shiliang Pu, and Zhan Ma, “High-efficiency lossy image coding through adaptive neighborhood information aggre- gation,” arXiv preprint arXiv:2204.11448 , 2022

    work page Pith review arXiv 2022

Show all 40 references
  1. [9]

    Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding,

    Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang, “Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June...

    2022

  2. [10]

    Learned image compression with mixed transformer-cnn architectures,

    Jinming Liu, Heming Sun, and Jiro Katto, “Learned image compression with mixed transformer-cnn architectures,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 14388–14397

    2023

  3. [11]

    High-fidelity generative image compression,

    Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson, “High-fidelity generative image compression,” in Advances in Neural Information Processing Systems . 2020, vol. 33, pp. 11913– 11924, Curran Associates, Inc

    2020

  4. [12]

    Improving statistical fidelity for neural image compression with implicit local likelihood models,

    Matthew J. Muckley, Alaaeldin El-Nouby, Karen Ullrich, Herve Jegou, and Jakob Verbeek, “Improving statistical fidelity for neural image compression with implicit local likelihood models,” in Proceedings of the 40th International Conference on Machine Learning . 23–29 Jul 2023,...

    2023

  5. [13]

    Variable-rate deep image compression through spatially-adaptive feature transform,

    Myungseo Song, Jinyoung Choi, and Bohyung Han, “Variable-rate deep image compression through spatially-adaptive feature transform,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2021, pp. 2380–2389

    2021

  6. [14]

    Transtic: Transferring transformer- based image compression from human perception to machine percep- tion,

    Yi-Hsin Chen, Ying-Chieh Weng, Chia-Hao Kao, Cheng Chien, Wei- Chen Chiu, and Wen-Hsiao Peng, “Transtic: Transferring transformer- based image compression from human perception to machine percep- tion,” in Proceedings of the IEEE/CVF International Conference on Computer Vision...

    2023

  7. [15]

    Image compression for machine and human vision with spatial-frequency adaptation,

    Han Li, Shaohui Li, Shuangrui Ding, Wenrui Dai, Maida Cao, Chenglin Li, Junni Zou, and Hongkai Xiong, “Image compression for machine and human vision with spatial-frequency adaptation,” in Computer Vision – ECCV 2024 , Cham, 2024, pp. 382–399, Springer Nature Switzerland

    2024

  8. [16]

    All-in-one image coding for joint human-machine vision with multi-path aggregation,

    Xu Zhang, Peiyao Guo, Ming Lu, and Zhan Ma, “All-in-one image coding for joint human-machine vision with multi-path aggregation,” in Advances in Neural Information Processing Systems . 2024, vol. 37, pp. 71465–71503, Curran Associates, Inc

    2024

  9. [17]

    Umap: Uniform manifold approximation and projection for dimension reduction,

    Leland McInnes, John Healy, and James Melville, “Umap: Uniform manifold approximation and projection for dimension reduction,” arXiv preprint arXiv:1802.03426, 2018

    2018 arXiv

  10. [18]

    Imagenet: A large-scale hierarchical image database,

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei- Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE Conference on Computer Vision and Pattern Recognition , 2009, pp. 248–255

    2009

  11. [19]

    Towards image understanding from deep compression without decoding,

    R ´obert Torfason, Fabian Mentzer, Eir ´ıkur ´Ag´ustsson, Michael Tschan- nen, Radu Timofte, and Luc Van Gool, “Towards image understanding from deep compression without decoding,” in International Conference on Learning Representations , 2018

    2018

  12. [20]

    Improving multiple machine vision tasks in the compressed domain,

    Jinming Liu, Heming Sun, and Jiro Katto, “Improving multiple machine vision tasks in the compressed domain,” in 2022 26th International Conference on Pattern Recognition (ICPR) , 2022, pp. 331–337

    2022

  13. [21]

    Unified architecture adaptation for compressed domain semantic inference,

    Zhihao Duan, Zhan Ma, and Fengqing Zhu, “Unified architecture adaptation for compressed domain semantic inference,” IEEE Trans. Circuit Syst. Video Technol. , vol. 33, no. 8, pp. 4108–4121, 2023

    2023

  14. [22]

    Image coding for machines with omnipotent feature learning,

    Ruoyu Feng, Xin Jin, Zongyu Guo, Runsen Feng, Yixin Gao, Tianyu He, Zhizheng Zhang, Simeng Sun, and Zhibo Chen, “Image coding for machines with omnipotent feature learning,” in Computer Vision – ECCV 2022 , Cham, 2022, pp. 510–528, Springer Nature Switzerland

    2022

  15. [23]

    Semantics- to-signal scalable image compression with learned revertible represen- tations,

    Kang Liu, Dong Liu, Li Li, Ning Yan, and Houqiang Li, “Semantics- to-signal scalable image compression with learned revertible represen- tations,” International Journal of Computer Vision , vol. 129, no. 9, pp. 2605–2621, 2021

    2021

  16. [24]

    Sssic: Semantics-to-signal scalable image coding with learned structural representations,

    Ning Yan, Changsheng Gao, Dong Liu, Houqiang Li, Li Li, and Feng Wu, “Sssic: Semantics-to-signal scalable image coding with learned structural representations,” IEEE Transactions on Image Processing , vol. 30, pp. 8939–8954, 2021

    2021

  17. [25]

    Scalable image coding for humans and machines,

    Hyomin Choi and Ivan V . Baji ´c, “Scalable image coding for humans and machines,” IEEE Transactions on Image Processing , vol. 31, pp. 2739–2754, 2022

    2022

  18. [26]

    Gen- erative adversarial nets,

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio, “Gen- erative adversarial nets,” in Advances in Neural Information Processing Systems. 2014, vol. 27, pp. 2672–2680, Curran Associates, Inc

    2014

  19. [27]

    The unreasonable effectiveness of deep features as a perceptual metric,

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2018, pp. 586–595

    2018

  20. [28]

    Deep residual learning for image recognition,

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2016, pp. 770–778

    2016

  21. [29]

    A convnet for the 2020s,

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie, “A convnet for the 2020s,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , June 2022, pp. 11976–11986

    2022

  22. [30]

    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 Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , October 2021, pp. 10012–10022

    2021

  23. [31]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weis- senborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby, “An image is worth 16x16 words: Transformers for image recognition...

    2021

  24. [32]

    Training data-efficient image transformers & distillation through attention,

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou, “Training data-efficient image transformers & distillation through attention,” in Proceedings of the 38th International Conference on Machine Learning . 18–24 Jul 2021, vol. ...

    2021

  25. [33]

    Kodak lossless true color image suite,

    Eastman Kodak, “Kodak lossless true color image suite,” 1993

    1993

  26. [34]

    Work- shop and challenge on learned image compression (clic2020),

    George Toderici, Wenzhe Shi, Radu Timofte, Lucas Theis, Johannes Balle, Eirikur Agustsson, Nick Johnston, and Fabian Mentzer, “Work- shop and challenge on learned image compression (clic2020),” 2020

    2020

  27. [35]

    Semantic understanding of scenes through the ade20k dataset,

    Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba, “Semantic understanding of scenes through the ade20k dataset,” International Journal of Computer Vision , vol. 127, pp. 302–321, 2019

    2019

  28. [36]

    A unified end-to- end framework for efficient deep image compression,

    Jiaheng Liu, Guo Lu, Zhihao Hu, and Dong Xu, “A unified end-to- end framework for efficient deep image compression,” arXiv preprint arXiv:2002.03370, 2020

    2002 arXiv

  29. [37]

    Ntire 2017 challenge on single image super-resolution: Dataset and study,

    Eirikur Agustsson and Radu Timofte, “Ntire 2017 challenge on single image super-resolution: Dataset and study,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, July 2017

    2017

  30. [38]

    Versatile video coding reference software version 17.1,

    “Versatile video coding reference software version 17.1,” https://vcgit. hhi.fraunhofer.de/jvet/VVCSoftware VTM/tags/VTM-17.1, July 2022

    2022

  31. [39]

    Pyramid scene parsing network,

    Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia, “Pyramid scene parsing network,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , July 2017, pp. 2881–2890

    2017

  32. [40]

    Adam: A method for stochastic optimization,

    Diederik P Kingma and Jimmy Ba, “Adam: A method for stochastic optimization,” in the 3rd Int. Conf. on Learning Representations , 2015

    2015

Pith tools

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