SegDINO: An Efficient Design for Medical and Natural Image Segmentation with DINO-V3

Hongqiu Wang; Lei Zhu; Sicheng Yang; Sixiang Chen; Zhaohu Xing

REVIEW 3 major objections 5 minor 3 cited by

SegDINO: An Efficient Design for Medical and Natural Image Segmentation with DINO-V3

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read SegDINO couples a frozen DINOv3-S backbone with a 2.21M-parameter MLP head and reports state-of-the-art segmentation on six medical and natural image benchmarks.

desk verdict First frozen-DINOv3 segmentation baseline, but the SOTA claim lacks documented fair baselines and any variance/ablation evidence. read the letter →

arxiv 2509.00833 v1 pith:7RJYH7OR submitted 2025-08-31 cs.CV

Sicheng Yang , Hongqiu Wang , Zhaohu Xing , Sixiang Chen , Lei Zhu This is my paper

classification cs.CV

keywords SegDINOimagesegmentationDINOv3frozenbackbonelightweightdecodermedicalmirrordetectionshadow

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 argues that the standard recipe for adapting self-supervised vision transformers to segmentation—heavy decoders with multi-scale fusion and complex upsampling—is unnecessary. SegDINO freezes a DINOv3-S backbone, gathers patch tokens from layers 3, 6, 9, and 12, aligns them to a common resolution and channel width, and lets a lightweight MLP head predict masks directly. The authors report that this system, with only 2.21 million trainable parameters, outperforms specialized baselines on three medical datasets (TN3K, Kvasir-SEG, ISIC) and three natural image datasets (MSD, VMD-D, ViSha). The claim matters because it identifies the pretrained encoder, not the decoder, as the main source of segmentation quality. If correct, it points toward a much cheaper route to high-accuracy segmentation in resource-constrained settings.

What carries the argument

The load-bearing component is the Light-Decoder (L-Decoder), a parameter-lean MLP head that consumes multi-level features from the frozen DINOv3-S encoder. Features from layers 3, 6, 9, and 12 are reformed—upsampled and channel-aligned following the strategy used in (Ranftl et al., 2021)—then concatenated along the channel dimension and mapped to per-pixel class logits. This single alignment-and-concatenation mechanism is what lets a 2.21M-parameter head replace much heavier decoders.

What would settle it

Re-run every baseline with its own native resolution and original training recipe on TN3K, Kvasir-SEG, ISIC, MSD, VMD-D, and ViSha, and compare Dice/IoU with SegDINO at 256x256; if SegDINO stops leading, the state-of-the-art claim rests on the common low-resolution protocol rather than on the model design itself.

Watch

Extended reading notes

Core claim

The paper's central discovery, stated in its own terms, is that a frozen DINOv3-S backbone requires almost no decoding machinery to segment. Multi-level patch tokens from four Transformer depths are upsampled, projected to a common channel width, concatenated, and passed through a small MLP that outputs class maps. With only the decoder trained, SegDINO reports the best Dice/IoU on TN3K, Kvasir-SEG, and ISIC, and the best IoU and F-measure on MSD, VMD-D, and ViSha, while running at 53 FPS. The authors take this as evidence that foundation features, properly aligned, carry enough dense semantic information that heavy decoder complexity becomes a cost rather than a benefit.

Load-bearing premise

The state-of-the-art claim assumes that comparing all models at 256x256 with cross-entropy loss is fair to baselines designed for higher resolutions and different loss schedules; if those baselines are handicapped, SegDINO's reported margins are inflated.

Editorial extensions

If this is right

Training SegDINO on a new dataset means training only the small head, so new segmentation tasks may require far less data and compute.
The recipe gives a concrete upper bound on the decoder complexity needed for competitive segmentation, shifting attention toward the quality of self-supervised encoders.
Because the encoder is frozen, one DINOv3 feature computation could serve multiple task-specific heads sharing the same backbone.
At 53 FPS with 2.21M trainable parameters, the approach is a candidate for clinical or edge deployment where model size and speed matter.

Reading between the lines

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

A natural test is replacing the fixed layer set {3, 6, 9, 12} with learned feature weighting or a searched layer subset, which could reveal how much of the gain depends on those particular depths.
The paper's fixed 256x256 protocol leaves open whether the reported margins persist at native resolutions; a fair-resolution rerun is the most direct stress test of the claim.
The frozen-encoder limitation the authors acknowledge could be addressed by adding small trainable adapters on top of DINOv3 for rare or highly domain-specific medical cases.
If the result generalizes, segmentation could become a 'train a head' problem rather than a 'design a decoder' problem—the paper's strongest implicit bet.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know two things: SegDINO is the first segmentation result using a frozen DINOv3 backbone with a 2.21M-parameter MLP head, and the numbers look strong across six benchmarks — but the evaluation protocol is under-specified and the central SOTA claim does not hold up as reported.

What is genuinely useful: the design is clean. Freeze DINOv3-S, take layers 3/6/9/12, upsample to a common resolution, concatenate, run an MLP. That is a direct DPT-style reformulation, so novelty is incremental, but it is the first reported use of DINOv3 for segmentation and the benchmark numbers are new. Code is released, and the efficiency story (2.21M trainable parameters, 53 FPS) is presented clearly.

The load-bearing problem is the baseline protocol. Section 3.2 says all images are resized to 256x256 "for consistent input resolution across models," but the paper never states whether the baselines (Mask2Former, SegFormer, VMD-Net, etc.) were retrained under the same settings or whether their numbers were quoted from original papers. Those methods generally rely on higher resolutions and specialized training objectives; at 256x256 with plain cross-entropy they are likely handicapped. Combined with no error bars or multiple seeds, the +5.4 IoU on MSD and +19.5 IoU on VMD-D could be protocol artifacts rather than genuine improvements. The paper itself admits in the conclusion that "further ablation studies are needed" — so the layer subset, decoder width, and resolution choices are unexamined. That is an explicit limitation and it weighs heavily.

That said, I do not think this is fabricated. The consistency across six diverse datasets is a real empirical signal that frozen DINOv3 features carry substantial segmentation information. But the paper overclaims by calling these results SOTA as reported. It is a plausible technical report with an incomplete evaluation.

The right audience is anyone deciding whether to build segmentation heads on DINOv3, plus methodology discussions about fair baseline comparisons. It deserves a serious referee because the question matters and the code makes the work reproducible, but I would not accept it in current form. I would ask for retrained baselines at matched resolution, error bars over at least three seeds, and a minimal ablation of the four design choices. If those come back, this could be a decent short paper or a strong baseline reference.

Referee Report

3 major / 5 minor

Summary. SegDINO couples a frozen DINOv3-S ViT backbone with a lightweight MLP decoder for image segmentation. Multi-level patch tokens from layers 3, 6, 9, and 12 are reformulated, concatenated, and passed to an MLP head that directly predicts the segmentation mask; only the decoder is trained. The method is evaluated on three medical datasets (TN3K, Kvasir-SEG, ISIC) and three natural-image datasets (MSD, VMD-D, ViSha). All inputs are resized to 256x256, training uses cross-entropy loss with AdamW for 50 epochs at batch size 4. The paper reports consistent improvements over strong baselines and claims state-of-the-art performance with only 2.21M trainable parameters.

Significance. If the empirical claims hold, the paper would make a useful point: a frozen self-supervised foundation model with a nearly trivial decoder can outperform task-specific architectures that are much larger and more complex. The strengths are the efficiency-oriented design, the breadth of benchmarks (medical, static mirror, video mirror, and video shadow), the explicit efficiency comparison, and the release of code. These factors make the contribution potentially reproducible and practically attractive. However, the central claim is purely empirical, and the evidence as presented is not yet sufficient: baseline evaluation protocol is under-specified, no multiple runs or statistical tests are reported, and no ablations substantiate the design choices. The paper explicitly acknowledges that 'further ablation studies are needed,' which supports the view that the central architecture claims are currently under-supported.

major comments (3)

[Section 3.2, Tables 1-4] The baseline evaluation protocol is the load-bearing assumption. The text says all images are resized to 256x256 and describes SegDINO's optimizer, loss, epochs, and batch size, but it never states whether each baseline was retrained under the same protocol or whether the numbers are quoted from original papers. This matters especially for methods like SegFormer and Mask2Former, which are typically evaluated at higher resolutions, and for video-based methods (TVSD, STICT, Scotch-Soda, VMD-Net) that exploit temporal or high-resolution cues. Forcing all inputs to 256x256 with cross-entropy loss can disproportionately harm those baselines. The reported margins are large (e.g., Table 3: +19.5 IoU over VMD-Net), so protocol artifacts could explain the results. The paper should specify per-baseline settings, ideally by retraining all baselines under the same preprocessing and training configur
[Tables 1-4, Section 3.2] No statistical reliability is reported. All tables contain single point estimates without error bars, multiple seeds, or significance tests. Some of the claimed gains are small (e.g., Table 4: SegDINO 0.675 vs. TBG-Diff 0.667 in IoU), while others are very large (Table 3: 0.762 vs. 0.567). Without variance estimates, the reader cannot tell which differences are meaningful. The authors should provide mean and standard deviation over at least three random seeds and, where feasible, paired significance tests for the key comparisons.
[Section 3.2 and Section 4] The central design choices are not ablated. The layer subset L={3,6,9,12}, the decoder width/depth, the 256x256 resolution, and the training recipe are fixed without supporting analysis. The conclusion itself states that 'further ablation studies are needed to better understand the contributions of individual components, such as feature selection depth, reformulation strategies, and decoder design.' Since the paper's contribution is specifically this frozen-backbone plus light-decoder design, the lack of ablations leaves its main technical claim unsubstantiated. At minimum, the layer subset and decoder size should be varied, and the 256x256 choice should be justified against higher resolutions.

minor comments (5)

[Section 3.3] The text says 'both variants achieve consistent improvements,' but only one SegDINO configuration is presented. This appears to be a leftover from an earlier draft and should be corrected.
[Section 2.3] The 'reformulated feature map' eZ is used in Eq. (4) but the reformulation operation is not defined. Please specify the upsampling/interpolation method, the channel projection, and the exact MLP architecture (number of layers, hidden widths, activations).
[Figure 2] The caption says 'Overall performance and efficiency comparisons across different datasets,' but the figure contains three separate panels. Please label each panel more clearly and define the plotted quantities in the caption.
[Section 3.3, Tables 2-4] For the natural-image tables, the dimensions reported are percentage-point differences, not relative percentages. The phrases 'over 2% in accuracy' and 'over 5% in F-beta' should be phrased as percentage points to avoid ambiguity.
[References] Some baseline names in Tables 2-4 are not explicitly associated with a table-side citation; please add references directly in the table captions or in the comparison text for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; empirical SOTA claim rests on external benchmarks, with protocol fairness as the main (non-circular) risk.

full rationale

SegDINO is an empirical systems paper, not a derivation. The method section (Eqs. 1–5) merely defines a frozen DINOv3 backbone, a subset of intermediate layers, channel-wise concatenation, and an MLP decoder; none of these equations purport to derive the reported accuracy, and the reported numbers come from external benchmarks (TN3K, Kvasir-SEG, ISIC, MSD, VMD-D, ViSha). No fitted parameter is renamed as a prediction: the only trainable component is the MLP head, and its weights are learned from the same train/test protocol used for the baselines. The paper does not invoke a uniqueness theorem, and the cited prior work on DINOv3 (Siméoni et al., 2025) and on upsampling/channel reformulation (Ranftl et al., 2021) is external, not self-referential. Some cited baselines (TVSD, Scotch-Soda, TBG-Diff) include the present authors, but they are used as comparison methods, not as load-bearing support for SegDINO's design. The paper itself acknowledges (Section 4) that ablations of feature-selection depth, reformulation, and decoder design are missing; that is a transparency gap, not circularity, because the central SOTA claim is not derived from those unablated choices. The main substantive risk is evaluation fairness: Section 3.2 states all images are resized to 256×256 and does not specify whether every baseline was retrained under identical settings. That is a protocol-fairness concern about the empirical comparison, not a case where the paper's conclusion is equivalent to its inputs by construction. Therefore no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim depends on several unstated design choices (layer subset, MLP dimensions, input resolution) and on assumptions about the transferability of DINOv3 features and the fairness of the evaluation protocol. No new physical or mathematical entities are introduced.

free parameters (4)

Layer subset L = {3, 6, 9, 12} = Not reported; chosen by hand
Sec 3.2 states these layers are extracted. No ablation justifies this selection; it is likely tuned on the same benchmarks used to claim SOTA.
Decoder MLP architecture (hidden width, depth, channel width C) = Not specified
Sec 2.3 describes an MLP decoder without giving C, number of layers, or hidden dimensions. These are free design choices that materially affect the reported accuracy.
Input resolution 256x256 = 256x256
Sec 3.2 sets all images to 256x256. This resolution choice interacts with every baseline and may unfairly disadvantage methods designed for higher resolutions.
Optimizer hyperparameters (lr, weight decay, epochs, batch size) = 1e-4, 1e-4, 50, 4
Sec 3.2 lists these without sensitivity analysis. The central claim depends on this single configuration.

assumptions (3)

domain assumption DINOv3 features are transferable to dense segmentation without fine-tuning.
The backbone is frozen in Sec 2.2. The paper assumes these features contain enough spatial and semantic information for segmentation without offering independent evidence beyond its own results.
domain assumption The DPT-style reformulation strategy transfers to DINOv3 features.
Sec 2.3 says it 'follows a reform strategy similar to' Ranftl et al., 2021. This transfer is assumed, not validated against alternative reformulations.
domain assumption The official train/test splits are used and the baseline numbers are comparable.
Sec 3.2 says the official splits are followed, but no details are given on how baselines were run or whether they share the same preprocessing, loss, and training budget.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SegDINO: An Efficient Design for Medical and Natural Image Segmentation with DINO-V3." pith.science (2026). https://pith.science/paper/7RJYH7OR

@misc{pith2026250900833,
  author       = {Pith},
  title        = {Pith review of: SegDINO: An Efficient Design for Medical and Natural Image Segmentation with DINO-V3},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7RJYH7OR}},
  note         = {Machine review of arXiv:2509.00833}
}

read the original abstract

The DINO family of self-supervised vision models has shown remarkable transferability, yet effectively adapting their representations for segmentation remains challenging. Existing approaches often rely on heavy decoders with multi-scale fusion or complex upsampling, which introduce substantial parameter overhead and computational cost. In this work, we propose SegDINO, an efficient segmentation framework that couples a frozen DINOv3 backbone with a lightweight decoder. SegDINO extracts multi-level features from the pretrained encoder, aligns them to a common resolution and channel width, and utilizes a lightweight MLP head to directly predict segmentation masks. This design minimizes trainable parameters while preserving the representational power of foundation features. Extensive experiments across six benchmarks, including three medical datasets (TN3K, Kvasir-SEG, ISIC) and three natural image datasets (MSD, VMD-D, ViSha), demonstrate that SegDINO consistently achieves state-of-the-art performance compared to existing methods. Code is available at https://github.com/script-Yang/SegDINO.

Figures

Figures reproduced from arXiv: 2509.00833 by the authors.

**Figure 2.** Overall performance and efficiency comparisons across different datasets. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Step-Attention Refinement of DINOv3 Features for Efficient Anterior Eye Segmentation
cs.CV 2026-07 conditional novelty 5.0 of 10

Step-attention refinement of multi-level DINOv3 features plus a light conv decoder yields 85.55% mIoU and best domain-shift robustness on seven-class clinical anterior-eye segmentation.
The pretraining domain outweighs the training objective in setting the privacy-utility trade-off of differentially private medical image analysis
cs.CV 2026-01 reject novelty 5.0 of 10

In DP-SGD chest X-ray classification, MIMIC-CXR supervised pretraining beats ImageNet and DINOv3 initializations, but the study cannot cleanly separate pretraining domain from objective because key comparison arms are...
LUMOS: Latent Universal Medical Priors for Segmentation
cs.CV 2026-03 reject novelty 4.0 of 10

A frozen vision model is used to build a ground-truth-trained guide mask that gates medical segmentation networks, but reported gains are inconsistent across datasets and the abstract and body describe different methods.

Reference graph

Works this paper leans on

14 extracted references · 3 canonical work pages · cited by 3 Pith papers

[1]

Recur- rent residual convolutional neural network based on u-net (r2u-net) for medical image segmenta- tion

Md Zahangir Alom, Mahmudul Hasan, Chris Yakopcic, Tarek M Taha, and Vijayan K Asari. Recur- rent residual convolutional neural network based on u-net (r2u-net) for medical image segmenta- tion. arXiv preprint arXiv:1802.06955,

arXiv
[4]

Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion anal- ysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isi...

work page 2017
[5]

Anomalydino: Boosting patch-based few-shot anomaly detection with dinov2

Simon Damm, Mike Laszkiewicz, Johannes Lederer, and Asja Fischer. Anomalydino: Boosting patch-based few-shot anomaly detection with dinov2. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 1319–1329. IEEE,

work page 2025
[10]

Dinov2: Learning robust visual features without supervision

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

arXiv
[11]

Oriane Sim´eoni, Huy V V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha ¨el Ramamonjisoa, et al. Dinov3. arXiv preprint arXiv:2508.10104,

arXiv
[14]

Timeline and boundary guided diffusion network for video shadow detection

Haipeng Zhou, Hongqiu Wang, Tian Ye, Zhaohu Xing, Jun Ma, Ping Li, Qiong Wang, and Lei Zhu. Timeline and boundary guided diffusion network for video shadow detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pp. 166–175, 2024a. Tianfei Zhou, Wang Xia, Fei Zhang, Boyu Chang, Wenguan Wang, Ye Yuan, Ender Konukoglu, and Daniel C...

arXiv
[2015]

Decoupled weight decay regularization

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101,

arXiv
[2017]

Hongqiu Wang, Jian Chen, Shichen Zhang, Yuan He, Jinfeng Xu, Mengwan Wu, Jinlan He, Wenjun Liao, and Xiangde Luo. Dual-reference source-free active domain adaptation for nasopharyngeal carcinoma tumor segmentation across multiple hospitals.IEEE Transactions on Medical Imaging, 43(12):4078–4090, 2024a. Hongqiu Wang, Guang Yang, Shichen Zhang, Jing Qin, Yik...

arXiv

Show all 14 references

[2018]

Segdiff: Image segmentation with diffusion probabilistic models

Tomer Amit, Tal Shaharbany, Eliya Nachmani, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390,

arXiv
[2019]

Attention u-net: Learning where to look for the pancreas

Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999,

arXiv
[2021]

Transunet: Transformers make strong encoders for medical image segmentation

Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021a. Wuyang Chen, Xianzhi Du, Fan Yang, Lucas Beyer, Xi...

arXiv
[2022]

U-mamba: Enhancing long-range dependency for biomedical image segmentation

Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722,

arXiv
[2023]

Re- visiting shadow detection: A new benchmark dataset for complex world

Xiaowei Hu, Tianyu Wang, Chi-Wing Fu, Yitong Jiang, Qiong Wang, and Pheng-Ann Heng. Re- visiting shadow detection: A new benchmark dataset for complex world. IEEE Transactions on Image Processing, 30:1925–1934,

1925
[2024]

Fast segment anything

10 Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything. arXiv preprint arXiv:2306.12156,

arXiv

Pith tools

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

[1] [1]

Recur- rent residual convolutional neural network based on u-net (r2u-net) for medical image segmenta- tion

Md Zahangir Alom, Mahmudul Hasan, Chris Yakopcic, Tarek M Taha, and Vijayan K Asari. Recur- rent residual convolutional neural network based on u-net (r2u-net) for medical image segmenta- tion. arXiv preprint arXiv:1802.06955,

arXiv

[2] [4]

Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion anal- ysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isi...

work page 2017

[3] [5]

Anomalydino: Boosting patch-based few-shot anomaly detection with dinov2

Simon Damm, Mike Laszkiewicz, Johannes Lederer, and Asja Fischer. Anomalydino: Boosting patch-based few-shot anomaly detection with dinov2. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 1319–1329. IEEE,

work page 2025

[4] [10]

Dinov2: Learning robust visual features without supervision

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

arXiv

[5] [11]

Oriane Sim´eoni, Huy V V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha ¨el Ramamonjisoa, et al. Dinov3. arXiv preprint arXiv:2508.10104,

arXiv

[6] [14]

Timeline and boundary guided diffusion network for video shadow detection

Haipeng Zhou, Hongqiu Wang, Tian Ye, Zhaohu Xing, Jun Ma, Ping Li, Qiong Wang, and Lei Zhu. Timeline and boundary guided diffusion network for video shadow detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pp. 166–175, 2024a. Tianfei Zhou, Wang Xia, Fei Zhang, Boyu Chang, Wenguan Wang, Ye Yuan, Ender Konukoglu, and Daniel C...

arXiv

[7] [2015]

Decoupled weight decay regularization

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101,

arXiv

[8] [2017]

Hongqiu Wang, Jian Chen, Shichen Zhang, Yuan He, Jinfeng Xu, Mengwan Wu, Jinlan He, Wenjun Liao, and Xiangde Luo. Dual-reference source-free active domain adaptation for nasopharyngeal carcinoma tumor segmentation across multiple hospitals.IEEE Transactions on Medical Imaging, 43(12):4078–4090, 2024a. Hongqiu Wang, Guang Yang, Shichen Zhang, Jing Qin, Yik...

arXiv