Pith. sign in

REVIEW 4 major objections 6 minor 63 references

Training a salient object detector solely on 139,981 synthetic images with masks decoded from diffusion and self-supervised features cuts cross-dataset error by 20–50%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 08:13 UTC pith:HTQEVAAF

load-bearing objection Solid synthetic-data result for SOD with a few presentation and label-quality questions worth fixing before publication. the 4 major comments →

arxiv 2510.21605 v3 pith:HTQEVAAF submitted 2025-10-24 cs.CV

S3OD: Towards Generalizable Salient Object Detection with Synthetic Data

Orest Kupyn , Hirokatsu Kataoka , Christian Rupprecht This is my paper
classification cs.CV
keywords synthetic datasalient object detectiondiffusion model featuresmulti-mask decoderambiguity-aware segmentationself-supervised featurescross-dataset generalizationhigh-resolution segmentation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that data scarcity, not model architecture, is the real bottleneck in salient object detection (SOD), and that synthetic data can replace expensive human pixel annotations. The authors generate S3OD, a dataset of 139,981 high-resolution images whose masks are read out of the same diffusion model that creates the images, fusing diffusion-transformer features, concept attention maps, and self-supervised visual features into pixel labels. Trained on synthetic data alone, their model reduces mean absolute error by 20–50% across DIS, HRSOD, UHRSD, DAVIS-S, DUTS, and DUT-OMRON compared to models trained on real datasets; fine-tuning then pushes it past prior state-of-the-art on DIS and HR-SOD. A reader should care because pixel-precise SOD annotations cost up to 10 hours per image, and if this label-generation pipeline works, it points a way to train dense-prediction models without manual annotation.

Core claim

The central discovery is that reliable pixel masks can be extracted for free during image generation by combining three complementary signals: feature maps from four layers of a FLUX diffusion transformer, concept attention maps that localize the object and background tokens, and DINO-v3 self-supervised features from the decoded image. Fusing these and decoding masks yields 139,000+ image-mask pairs in three feedback-driven rounds, with a multi-stage filter (consistency check and vision-language validation) removing 6.8% of samples. A model trained only on this synthetic data outperforms models trained on real datasets in cross-dataset tests, and after task-specific fine-tuning it reaches st

What carries the argument

The load-bearing mechanism is the multi-modal diffusion label-extraction pipeline: FLUX DiT feature maps, concept attention maps (object vs background), and DINO-v3 visual features are each projected, fused, and decoded into masks, so labels come from the generator itself rather than from a teacher network. The other central object is the ambiguity-aware decoder: a DPT-style model with DINO-v3 backbone that outputs several masks plus per-mask IoU estimates, trained with multiple-choice learning and a decaying auxiliary loss, then selects the highest-IoU prediction at inference. This design makes synthetic labels cheap and lets the model represent multiple valid interpretations of saliency.

Load-bearing premise

The entire result rests on the assumption that the masks decoded from diffusion features, concept attention, and DINO-v3 agree with what humans would mark as salient; the paper inspects only 1,000 samples and filters by consistency and a vision-language model rather than measuring against human ground truth at scale.

What would settle it

Take a new set of 5,000 real images with human pixel annotations from object categories not in S3OD's 1,676 classes, train the same model on S3OD alone, and compare its MAE on that set against a model trained on DUTS/DIS-5K; if the synthetic-only model does not reproduce a 20–50% error reduction, the headline generalization claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A model trained exclusively on S3OD generalizes to real-world benchmarks without any real training images, reducing MAE by 20–50% relative to models trained on conventional SOD datasets.
  • Fine-tuning S3OD-pretrained models on DIS-5K and HR-SOD datasets achieves state-of-the-art results, with reported error-rate reductions of 7–27% over the previous best on DIS-5K and DUT-OMRON.
  • The iterative generation loop—re-weighting categories by the current model's errors—improves F-measure by 3.6% on DIS and 5.3% on DUT-OMRON, showing closed-loop data synthesis helps.
  • All three label sources (DiT maps, concept attention, DINO-v3) are necessary; ablations show removing any one degrades cross-dataset performance.
  • Because the pipeline needs no external teacher mask predictor, it scales past 139,981 images and extends to other dense prediction tasks such as depth or edge detection.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves untested how far the 20–50% gain extends to domains far from FLUX's photorealistic training distribution (e.g., indoor clutter, documents, aerial views); an editor's guess is that its generalization is strongest for natural-scene imagery.
  • If the multi-mask decoder's oracle gap is any guide, existing saliency benchmarks may contain more annotator disagreement than is usually acknowledged; measuring inter-annotator IoU on a subset of DIS/DUTS would tell whether the ambiguity-aware design is a genuine fix or an artifact of the evaluation.
  • The iterative category re-weighting is a general recipe for closed-loop synthetic data: any task with a measurable per-category error could use the same 'sample more where the model fails' rule, though the paper only demonstrates it for SOD.
  • A cheaper way to stress-test the label pipeline would be to compare S3OD masks against human annotations on a random sample of generated images; if disagreement is high, the reported generalization stems from the model architecture or data scale, not from label fidelity.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces S3OD, a 139,981-image synthetic dataset for salient object detection, with masks produced by fusing FLUX-DiT feature maps, concept attention maps, and DINO-v3 features, followed by consistency and VLM filtering. It pairs this dataset with a DPT-style architecture using a DINO-v3 backbone and a multi-mask decoder that outputs several candidate masks plus estimated IoU scores for test-time selection. An iterative generation loop reweights object categories according to the student model's performance. The paper claims that models trained only on synthetic data achieve 20-50% error reduction in cross-dataset generalization (Table 2), and that fine-tuned variants reach state-of-the-art results on DIS and HR-SOD benchmarks (Table 3).

Significance. If the claims hold, the contribution is substantial: a scalable, teacher-free pathway to pixel-level supervision for dense prediction, an explicit treatment of label ambiguity in SOD, and a large public dataset. The final evaluation on external real benchmarks is appropriate, so the headline result is not tautologically defined. The ablations cover data modalities, backbone choice, number of mask branches, and number of generation rounds, which is a strength. However, the central assumption that automatically generated masks approximate human saliency is not quantitatively validated, and several reporting ambiguities affect reproducibility. The significance is therefore conditional on resolving the label-validity question and clarifying the training/evaluation procedures.

major comments (4)
  1. [§4.1, §4.3, §B] The abstract's load-bearing claim ('20-50% error reduction') assumes the 139,981 pseudo-masks faithfully approximate human saliency. The evidence provided is a self-audit of 1,000 samples (§B: 1.4% boundary issues, 0.1% missing objects) and filters that remove only 6.8% of samples (§4.3). This is insufficient to rule out systematic class- or scene-level label bias. I request: (i) a quantitative comparison of generated masks against human annotations on existing SOD test sets (e.g., DUTS-TE, MSRA-B, HRSOD) with per-category error breakdowns; (ii) an analysis of how the shared DINO-v3 representation between label extraction and student backbone affects the results, e.g., training a non-DINO student on the same pseudo-labels; (iii) pseudo-label quality statistics per generation round. Without this, the Table 2 gains may reflect the prior encoded in the label-generation pipeline rather than
  2. [§3.1, Figure 2] The multi-mask training procedure is described inconsistently. The text defines the best branch as i* = arg min IoU(m_i, y), which should be arg max since IoU is a similarity score. The text says the training loss uses the branch with the highest IoU with the ground-truth mask, while the Figure 2 caption says the loss is propagated through the branch with the highest predicted IoU. Test-time selection uses the predicted IoU head. Please clarify whether training uses the actual or the predicted IoU for branch selection, correct the arg min/arg max typo, and align the caption with the text. This is central to reproducing the multi-mask decoder.
  3. [§4.2, Table 5] The iterative generation framework is not described precisely enough to support the claim that feedback-driven reweighting helps. The held-out test set used for category evaluation is unspecified: which images form it, how is it constructed, and is it disjoint from the student model's training data and from the final evaluation benchmarks? Moreover, Table 5 compares one, two, and three rounds with increasing data size, so the observed gains could be due to data scale rather than the reweighting mechanism. Please add a matched-size control (e.g., randomly sampling the same number of new images) to isolate the effect of prioritizing hard categories.
  4. [§5.3, §5.5, Tables 3 and 5] Several quantitative claims in the text are not reproduced by the tables. §5.3 reports error-rate reductions of 14.0%, 7.3%, 20.6%, and 17.1%; §5.5 reports a 3.6% F-measure gain on DIS and 5.3% on DUT-OMRON from iterative generation. From Table 5, the overall DIS F-measure improves from 0.879 to 0.881 (about 0.2% relative) and DUT-OMRON from 0.855 to 0.860 (about 0.6% relative). Please state the exact definition of error rate used and recalculate; otherwise these headline improvements are unverifiable.
minor comments (6)
  1. [Table 2] The row labels 'S3OD SOD' and 'S3OD S3OD' are not defined in the text or caption. Based on §5.2 they seem to mean the S3OD architecture trained on SOD datasets and on synthetic S3OD data, respectively, but this should be explicit.
  2. [§3.2] In the final objective, the regularization term is written as λ_reg e^{-γt} L_mask(m_i) without a summation or index; it should be a sum over all branches i=1..N to match the surrounding text.
  3. [§B] For the manual audit of 1,000 samples, please report how the sample was stratified (categories, iterations, complexity) and provide a confidence interval; a single 1.4%/0.1% estimate has very wide uncertainty for rare failure modes.
  4. [Table 3] The S3OD* results are oracle metrics computed with ground-truth mask selection and are correctly said to be non-comparable, but they are bolded in the table. Please separate them visually from fair comparisons to avoid implying they are a state-of-the-art result.
  5. [Table 1, §4.4] The text states S3OD is '131% more' than the 11 most common academic benchmarks combined. Summing the datasets listed in Table 1 gives about 39k images, so 139,981 is roughly 258% more. Please verify the arithmetic or clarify which benchmarks are included.
  6. [General] There are minor typos, e.g., 'siginificantly' in §5.3 and the inconsistent use of DINO-v3 vs DINOv3 naming. A careful proofread is recommended.

Circularity Check

0 steps flagged

No significant circularity: the reported results are validated against external real benchmarks, and no prediction reduces to the pipeline's own definitions or to a load-bearing self-citation.

full rationale

The paper's headline numbers (Tables 2 and 3) are measured against human ground-truth masks of DIS-5K, HRSOD, UHRSD, DUTS-TE, and DUT-OMRON; none of these targets is produced by the S3OD pipeline. The pseudo-label generation described in §4.1 combines FLUX DiT features, concept attention maps, and DINO-v3 features, but the final evaluation does not use these pseudo-labels as the ground truth. DINO-v3 is used both as a label modality and as the student encoder (§3.1, §4.1), which is a legitimate risk of learning DINO-v3's saliency bias rather than human saliency; however, this is a generalization/correctness concern, not a circular derivation, because the benchmark annotations are independent of DINO-v3. The iterative category reweighting in §4.2 uses the model's own scores on an unspecified held-out test set to resample the synthetic distribution; even if this is a within-pipeline feedback loop, it does not determine the external benchmark numbers. The oracle S3OD* row in §5.3 is explicitly labeled as using ground truth and 'cannot be compared to other methods,' so it is not presented as a prediction. The manual 1,000-image audit in Appendix B and the VLM/consistency filters in §4.3 are quality checks, not evaluation targets. Self-citations (Kupyn & Rupprecht 2024; Kupyn et al. 2024) appear only in related-work context and are not load-bearing. I could not exhibit a specific reduction, via the paper's equations or definitions, from the pipeline's outputs to the reported benchmark results.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The paper introduces no new physical entities; the S3OD dataset is an artifact, not a postulated entity. All free parameters are training/filtering constants; the central claim rests on domain assumptions about diffusion/synthetic labels and pretrained features rather than on new mathematical structure.

free parameters (6)
  • Loss weights (λ_mask=10, λ_score=0.05, λ_reg=0.1, γ=0.2) = 10, 0.05, 0.1, 0.2
    Hand-chosen constants in §3.2; no sensitivity analysis, likely tuned on validation.
  • Number of prediction branches N = 3
    Ablated in Table 7; performance peaks at N=3.
  • Consistency filter threshold τ and mask heuristics = τ=0.8; ≤5 components; >70% coverage
    Multi-stage filter thresholds in §4.3 are hand-set and remove 6.8% of samples.
  • Iterative weighting constants α=8, β=0.5, w_min=1/|C|, w_new=4/|C| = 8, 0.5, 1/|C|, 4/|C|
    Non-linear category reweighting in §4.2; constants chosen to control feedback strength.
  • Focal loss focusing parameter τ = 2
    Standard focal loss setting, §3.2.
  • Feature projection dim / FLUX layers = 256; layers {4,16,27,36}
    Architecture choices for multi-modal fusion in §4.1; no ablation reported for these.
axioms (6)
  • domain assumption FLUX-Krea generates photorealistic, diverse scenes that approximate real-world saliency distributions.
    Central to 'synthetic data replaces real data'; supported only by FID/IS and a 1k-sample manual check.
  • domain assumption DINO-v3 semantic features align with human foreground/background judgments on generated images.
    Used both to create masks and as the student backbone; no independent per-pixel human validation for most samples.
  • domain assumption Concept attention maps from FLUX provide consistent object localization, and the background token complements them.
    Section 4.1 assumes two concepts (object category, background) suffice to decode masks.
  • domain assumption External benchmarks (DIS-5K, HRSOD, UHRSD, DUTS, DUT-OMRON, DAVIS-S) provide valid ground truth for comparing generalization.
    Evaluation assumptions; the paper accepts these as the external measure of saliency.
  • standard math Multiple-choice learning with relaxed assignment converges to a useful ambiguity-aware decoder.
    Borrowed from Guzman-Rivera 2012 and Rupprecht 2017; assumed to converge for this architecture.
  • domain assumption Image augmentations plus FLUX-Krea photorealism sufficiently reduce the synthetic-to-real domain gap.
    Mentioned in §B without quantitative evidence connecting directly to SOD performance.

pith-pipeline@v1.3.0-alltime-deepseek · 15249 in / 14960 out tokens · 129758 ms · 2026-08-04T08:13:55.338309+00:00 · methodology

0 comments
read the original abstract

Salient object detection exemplifies data-bounded tasks where expensive pixel-precise annotations force separate model training for related subtasks like DIS and HR-SOD. We present a method that dramatically improves generalization through large-scale synthetic data generation and ambiguity-aware architecture. We introduce S3OD, a dataset of over 139,000 high-resolution images created through our multi-modal diffusion pipeline that extracts labels from diffusion and DINO-v3 features. The iterative generation framework prioritizes challenging categories based on model performance. We propose a streamlined multi-mask decoder that handles the inherent ambiguity in salient object detection by predicting multiple valid interpretations. Models trained only on synthetic data achieve 20-50% error reduction in cross-dataset generalization, while fine-tuned versions reach state-of-the-art performance across DIS and HR-SOD benchmarks.

Figures

Figures reproduced from arXiv: 2510.21605 by Christian Rupprecht, Hirokatsu Kataoka, Orest Kupyn.

Figure 1
Figure 1. Figure 1: S3OD Top: Our large scale synthetic dataset, consisting of diverse complex scenes and high quality samples. Bottom: Model Predictions. Our model trained on synthetic data generalizes well to real-world images, handling ambiguous scenes by predicting alternative hypothesis. 1 INTRODUCTION Salient object detection (SOD) is a fundamental computer vision problem with applications spanning AR/VR Tian et al. (20… view at source ↗
Figure 2
Figure 2. Figure 2: S3OD Architecture. Model extends DPT Ranftl et al. (2021) to predict multiple mask candidates and a vector of IoUs with the ground truth, employing DINO-v3 as the backbone. During training, the loss is propagated through the branch with the highest predicted IoU. Focal Loss. To address foreground-background imbalance, we implement focal loss, widely used in dense prediction: Lfocal(mi) = − H X×W p=1 (1 − m… view at source ↗
Figure 3
Figure 3. Figure 3: Iterative Generation Pipeline. The LLM Achiam et al. (2023) generates a library of diverse prompts for a large set of object categories. These prompts guide a diffusion model to generate synthetic images with corresponding masks. The resulting dataset trains SOD model, which undergoes category-wise evaluation. Performance feedback from a trained SOD model dynamically adjusts category weights w˜i , prioriti… view at source ↗
Figure 4
Figure 4. Figure 4: S3OD Dataset: The dataset consists of diverse object categories and complex scenes that closely reflect real-world environments, featuring various lighting conditions, spatial compositions, and object interactions. All samples are generated with multi-modal dataset diffusion. images in category ci . The category weights w (r+1) i for the next iteration are updated proportionally to the inverse of these sco… view at source ↗
Figure 5
Figure 5. Figure 5: The complete system prompt used to instruct the LLM Achiam et al. (2023) for generating [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: S3OD Dataset Samples: Our method generates diverse high quality samples across a wide variety of object categories. object detection models. As shown in [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Prompt Enhancement: Top: Class name as a prompt. Bottom: LLM Prompt Generator. By focusing on key properties of salient object detection dataset the agent creates detailed and diverse prompts to maximize the diversity and realism. Input Image InSPyReNet BiRefNet MVANet S3OD Ground Truth [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison vs state-of-the-art methods on DIS-5K Qin et al. (2022) dataset. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

63 extracted references · 14 linked inside Pith

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

  3. [3]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  4. [4]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  5. [5]

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

  6. [6]

    Frequency-tuned salient region detection

    Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In 2009 IEEE conference on computer vision and pattern recognition, pp.\ 1597--1604. IEEE, 2009

    2009

  7. [7]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  8. [8]

    Synthetic data from diffusion models improves imagenet classification

    Shekoofeh Azizi, Simon Kornblith, Chitwan Saharia, Mohammad Norouzi, and David J Fleet. Synthetic data from diffusion models improves imagenet classification. arXiv preprint arXiv:2304.08466, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Flux.1 [krea-dev]: Photorealistic image generation with enhanced realism

    Black Forest Labs . Flux.1 [krea-dev]: Photorealistic image generation with enhanced realism. https://bfl.ai/blog/flux-1-krea-dev, 2025

    2025

  10. [10]

    Salient object detection: A benchmark

    Ali Borji, Ming-Ming Cheng, Huaizu Jiang, and Jia Li. Salient object detection: A benchmark. IEEE transactions on image processing, 24 0 (12): 0 5706--5722, 2015

    2015

  11. [11]

    Unseen salient object discovery for monocular robot vision

    Darren M Chan and Laurel D Riek. Unseen salient object discovery for monocular robot vision. IEEE Robotics and Automation Letters, 5 0 (2): 0 1484--1491, 2020

    2020

  12. [12]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  13. [13]

    Structure-measure: A new way to evaluate foreground maps

    Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE international conference on computer vision, pp.\ 4548--4557, 2017

    2017

  14. [14]

    Enhanced-alignment measure for binary foreground map evaluation

    Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, Ming-Ming Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421, 2018

    Pith/arXiv arXiv 2018

  15. [15]

    Scaling laws of synthetic images for model training

    Lijie Fan, Kaifeng Chen, Dilip Krishnan, Dina Katabi, Phillip Isola, and Yonglong Tian. Scaling laws of synthetic images for model training... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 7382--7392, 2024

    2024

  16. [16]

    Attentive feedback network for boundary-aware salient object detection

    Mengyang Feng, Huchuan Lu, and Errui Ding. Attentive feedback network for boundary-aware salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 1623--1632, 2019

    2019

  17. [17]

    Context-aware saliency detection

    Stas Goferman, Lihi Zelnik-Manor, and Ayellet Tal. Context-aware saliency detection. IEEE transactions on pattern analysis and machine intelligence, 34 0 (10): 0 1915--1926, 2011

    1915

  18. [18]

    Multiple choice learning: Learning to produce multiple structured outputs

    Abner Guzman-Rivera, Dhruv Batra, and Pushmeet Kohli. Multiple choice learning: Learning to produce multiple structured outputs. Advances in neural information processing systems, 25, 2012

    2012

  19. [19]

    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, pp.\ 770--778, 2016

    2016

  20. [20]

    Conceptattention: Diffusion transformers learn highly interpretable features, 2025

    Alec Helbling, Tuna Han Salih Meral, Ben Hoover, Pinar Yanardag, and Duen Horng Chau. Conceptattention: Diffusion transformers learn highly interpretable features, 2025. URL https://arxiv.org/abs/2502.04320

    Pith/arXiv arXiv 2025

  21. [21]

    Diffusion models for open-vocabulary segmentation

    Laurynas Karazija, Iro Laina, Andrea Vedaldi, and Christian Rupprecht. Diffusion models for open-vocabulary segmentation. In European Conference on Computer Vision, pp.\ 299--317. Springer, 2024

    2024

  22. [22]

    Segment anything in high quality

    Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36: 0 29914--29934, 2023

    2023

  23. [23]

    Revisiting image pyramid structure for high resolution salient object detection

    Taehun Kim, Kunhee Kim, Joonyeong Lee, Dongmin Cha, Jiho Lee, and Daijin Kim. Revisiting image pyramid structure for high resolution salient object detection. In Proceedings of the Asian Conference on Computer Vision, pp.\ 108--124, 2022

    2022

  24. [24]

    Dataset enhancement with instance-level augmentations

    Orest Kupyn and Christian Rupprecht. Dataset enhancement with instance-level augmentations. In European Conference on Computer Vision, pp.\ 384--402. Springer, 2024

    2024

  25. [25]

    Vggheads: A large-scale synthetic dataset for 3d human heads

    Orest Kupyn, Eugene Khvedchenia, and Christian Rupprecht. Vggheads: A large-scale synthetic dataset for 3d human heads. arXiv preprint arXiv:2407.18245, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2023. URL https://github.com/black-forest-labs/flux

    2023

  27. [27]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll \'a r. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pp.\ 2980--2988, 2017

    2017

  28. [28]

    Fully understanding generic objects: Modeling, segmentation, and reconstruction

    Feng Liu, Luan Tran, and Xiaoming Liu. Fully understanding generic objects: Modeling, segmentation, and reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 7423--7433, 2021 a

    2021

  29. [29]

    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, pp.\ 10012--10022, 2021 b

    2021

  30. [30]

    a henb \

    Federico Perazzi, Philipp Kr \"a henb \"u hl, Yael Pritch, and Alexander Hornung. Saliency filters: Contrast based filtering for salient region detection. In 2012 IEEE conference on computer vision and pattern recognition, pp.\ 733--740. IEEE, 2012

    2012

  31. [31]

    The 2017 davis challenge on video object segmentation

    Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbel \'a ez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017

    Pith/arXiv arXiv 2017

  32. [32]

    Maskfactory: Towards high-quality synthetic data generation for dichotomous image segmentation

    Haotian Qian, Yinda Chen, Shengtao Lou, Fahad Khan, Xiaogang Jin, and Deng-Ping Fan. Maskfactory: Towards high-quality synthetic data generation for dichotomous image segmentation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  33. [33]

    Basnet: Boundary-aware salient object detection

    Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan, and Martin Jagersand. Basnet: Boundary-aware salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 7479--7489, 2019

    2019

  34. [34]

    U2-net: Going deeper with nested u-structure for salient object detection

    Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern recognition, 106: 0 107404, 2020

    2020

  35. [35]

    Highly accurate dichotomous image segmentation

    Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In European Conference on Computer Vision, pp.\ 38--56. Springer, 2022

    2022

  36. [36]

    Vision transformers for dense prediction

    Ren \'e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 12179--12188, 2021

    2021

  37. [37]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R \"a dle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024

    Pith/arXiv arXiv 2024

  38. [38]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

    2022

  39. [39]

    Learning in an uncertain world: Representing ambiguity through multiple hypotheses

    Christian Rupprecht, Iro Laina, Robert DiPietro, Maximilian Baust, Federico Tombari, Nassir Navab, and Gregory D Hager. Learning in an uncertain world: Representing ambiguity through multiple hypotheses. In Proceedings of the IEEE international conference on computer vision, pp.\ 3591--3600, 2017

    2017

  40. [40]

    Fake it till you make it: Learning transferable representations from synthetic imagenet clones

    Mert B \"u lent Sar y ld z, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning transferable representations from synthetic imagenet clones. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8011--8021, 2023

    2023

  41. [41]

    Laion-5b: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35: 0 25278--25294, 2022

    2022

  42. [42]

    Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion

    Jordan Shipard, Arnold Wiliem, Kien Nguyen Thanh, Wei Xiang, and Clinton Fookes. Diversity is definitely needed: Improving model-agnostic zero-shot classification via stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 769--778, 2023

    2023

  43. [43]

    Oriane Sim \'e oni, Huy V Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha \"e l Ramamonjisoa, et al. Dinov3. arXiv preprint arXiv:2508.10104, 2025

    Pith/arXiv arXiv 2025

  44. [44]

    Gemma 3 technical report

    Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ram \'e , Morgane Rivi \`e re, et al. Gemma 3 technical report. arXiv preprint arXiv:2503.19786, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    Kine-appendage: Enhancing freehand vr interaction through transformations of virtual appendages

    Yang Tian, Hualong Bai, Shengdong Zhao, Chi-Wing Fu, Chun Yu, Haozhao Qin, Qiong Wang, and Pheng-Ann Heng. Kine-appendage: Enhancing freehand vr interaction through transformations of virtual appendages. IEEE Transactions on Visualization and Computer Graphics, 2022

    2022

  46. [46]

    Stablerep: Synthetic images from text-to-image models make strong visual representation learners

    Yonglong Tian, Lijie Fan, Phillip Isola, Huiwen Chang, and Dilip Krishnan. Stablerep: Synthetic images from text-to-image models make strong visual representation learners. arXiv preprint arXiv:2306.00984, 2023

    Pith/arXiv arXiv 2023

  47. [47]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  48. [48]

    Learning to detect salient objects with image-level supervision

    Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 136--145, 2017

    2017

  49. [49]

    Salient object detection with pyramid attention and salient edges

    Wenguan Wang, Shuyang Zhao, Jianbing Shen, Steven CH Hoi, and Ali Borji. Salient object detection with pyramid attention and salient edges. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 1448--1457, 2019

    2019

  50. [50]

    Label decoupling framework for salient object detection

    Jun Wei, Shuhui Wang, Zhe Wu, Chi Su, Qingming Huang, and Qi Tian. Label decoupling framework for salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 13025--13034, 2020

    2020

  51. [51]

    Datasetdm: Synthesizing data with perception annotations using diffusion models

    Weijia Wu, Yuzhong Zhao, Hao Chen, Yuchao Gu, Rui Zhao, Yefei He, Hong Zhou, Mike Zheng Shou, and Chunhua Shen. Datasetdm: Synthesizing data with perception annotations using diffusion models. arXiv preprint arXiv:2308.06160, 2023 a

    Pith/arXiv arXiv 2023

  52. [52]

    Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models

    Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. arXiv preprint arXiv:2303.11681, 2023 b

    Pith/arXiv arXiv 2023

  53. [53]

    Cascaded partial decoder for fast and accurate salient object detection

    Zhe Wu, Li Su, and Qingming Huang. Cascaded partial decoder for fast and accurate salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 3907--3916, 2019 a

    2019

  54. [54]

    Stacked cross refinement network for edge-aware salient object detection

    Zhe Wu, Li Su, and Qingming Huang. Stacked cross refinement network for edge-aware salient object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 7264--7273, 2019 b

    2019

  55. [55]

    Synthetic data supervised salient object detection

    Zhenyu Wu, Lin Wang, Wei Wang, Tengfei Shi, Chenglizhao Chen, Aimin Hao, and Shuo Li. Synthetic data supervised salient object detection. In Proceedings of the 30th ACM International Conference on Multimedia, pp.\ 5557--5565, 2022

    2022

  56. [56]

    Pyramid grafting network for one-stage high resolution saliency detection

    Chenxi Xie, Changqun Xia, Mingcan Ma, Zhirui Zhao, Xiaowu Chen, and Jia Li. Pyramid grafting network for one-stage high resolution saliency detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 11717--11726, 2022

    2022

  57. [57]

    Saliency detection via graph-based manifold ranking

    Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 3166--3173, 2013

    2013

  58. [58]

    Exploring limits of diffusion-synthetic training with weakly supervised semantic segmentation

    Ryota Yoshihashi, Yuya Otsuka, Tomohiro Tanaka, Hirokatsu Kataoka, et al. Exploring limits of diffusion-synthetic training with weakly supervised semantic segmentation. In Proceedings of the Asian Conference on Computer Vision, pp.\ 2300--2318, 2024

    2024

  59. [59]

    Multi-view aggregation network for dichotomous image segmentation

    Qian Yu, Xiaoqi Zhao, Youwei Pang, Lihe Zhang, and Huchuan Lu. Multi-view aggregation network for dichotomous image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 3921--3930, 2024

    2024

  60. [60]

    Towards high-resolution salient object detection

    Yi Zeng, Pingping Zhang, Jianming Zhang, Zhe Lin, and Huchuan Lu. Towards high-resolution salient object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 7234--7243, 2019

    2019

  61. [61]

    Progressive attention guided recurrent network for salient object detection

    Xiaoning Zhang, Tiantian Wang, Jinqing Qi, Huchuan Lu, and Gang Wang. Progressive attention guided recurrent network for salient object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 714--722, 2018

    2018

  62. [62]

    Egnet: Edge guidance network for salient object detection

    Jia-Xing Zhao, Jiang-Jiang Liu, Deng-Ping Fan, Yang Cao, Jufeng Yang, and Ming-Ming Cheng. Egnet: Edge guidance network for salient object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 8779--8788, 2019

    2019

  63. [63]

    Bilateral reference for high-resolution dichotomous image segmentation

    Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. arXiv preprint arXiv:2401.03407, 2024

    Pith/arXiv arXiv 2024