Pith. sign in

REVIEW 2 major objections 5 minor 34 references

A three-stage Generate-Filter-Refine pipeline lets SAM3 do referring camouflaged segmentation without any training, matching supervised specialists on R2C7K.

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 · grok-4.5

2026-07-14 03:33 UTC pith:K5F26HTX

load-bearing objection Solid training-free SOTA on R2C7K by rewiring SAM3 for cross-image exemplars; scoped but clean empirical work. the 2 major comments →

arxiv 2607.11732 v1 pith:K5F26HTX submitted 2026-07-13 cs.CV

GFR-SAM: Training-Free Referring Camouflaged Object Segmentation via Cross-Image Prompting

Yilong Yang , Jianxin Tian , Shengchuan Zhang , Liujuan Cao This is my paper
classification cs.CV
keywords Ref-CODTraining-FreeSAMIn-Context LearningCamouflaged Object DetectionCross-Image PromptingGenerate-Filter-Refine
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.

Referring Camouflaged Object Detection asks a model to find a hidden target that matches a reference cue, but supervised systems need dense labels and earlier training-free systems break when a single point prompt is slightly wrong. This paper claims that the same foundation models already contain the needed capability once the problem is recast as Generate-Filter-Refine. First, a reference image is turned into a clean visual prototype (mask-gated, position-free) that drives SAM3 across images to produce candidate masks. Second, a DINOv3 region-versus-global contrast score ranks those candidates and discards background distractors. Third, the top box plus a category name is fed back into SAM3 to recover fine boundaries and missing instances. On the R2C7K benchmark the resulting system raises weighted F-measure by 8.7 points over prior training-free methods and reaches scores that rival fully supervised specialists, all without updating a single parameter.

Core claim

GFR-SAM shows that SAM3 can be unlocked for cross-image referring camouflaged segmentation by a training-free Generate-Filter-Refine pipeline: mask-gated exemplar encoding produces candidates, DINOv3 region-global contrast selects the best one, and geometric-semantic co-prompting recovers complete instances, yielding state-of-the-art training-free numbers that compete with supervised models on R2C7K.

What carries the argument

The Generate-Filter-Refine pipeline, whose load-bearing step is In-Context Exemplar-guided Segmentation: Mask-guided Spatial Gating removes background from the reference RoI, coordinate embeddings are discarded, and the resulting pure visual prototype is used as a cross-image prompt for SAM3.

Load-bearing premise

That a fixed DINOv3 contrast score and a single confidence threshold of 0.35 will consistently surface the true camouflaged instance once the reference has been mask-gated, even when contrast is extreme or distractors are highly salient.

What would settle it

Run the identical pipeline with the same τ=0.35 and DINOv3 backbone on a new multi-instance, low-contrast camouflage set outside R2C7K; if weighted F-measure falls back to the level of plain point-prompt baselines, the claim that the Generate-Filter-Refine unlock is general fails.

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

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

2 major / 5 minor

Summary. The paper proposes GFR-SAM, a three-stage training-free pipeline for Referring Camouflaged Object Detection that adapts SAM3 for cross-image use. Ice-Seg builds a background-cleaned, position-invariant visual prototype from a reference mask via Mask-guided Spatial Gating (Eqs. 5–6) and feeds it to the SAM3 decoder to produce candidate masks; RGCF ranks them with a DINOv3 region-global contrast score (Eqs. 7–11); GSR then re-prompts SAM3 with the top-1 box plus a category text prompt at a low confidence threshold τ. On the R2C7K test set with N=3 references, the method reports overall S_m=0.894, αE=0.950, F_β^w=0.865, M=0.021, outperforming prior training-free baselines by a large margin and matching or exceeding several fully supervised Ref-COD specialists (Table 1). Ablations isolate MSG, Top-K, encoder choice, τ, prompting strategy, and shot count; qualitative failure modes are disclosed.

Significance. If the R2C7K numbers hold under independent re-implementation, the work supplies a practical, annotation-free alternative to supervised Ref-COD that is competitive with recent fully supervised specialists while remaining fully training-free. The concrete engineering contribution—unlocking SAM3’s exemplar pathway for true cross-image in-context prompting via MSG and coordinate discarding—is clearly articulated and ablated, and the Generate-Filter-Refine decomposition is a reusable pattern for other referring segmentation settings that currently rely on fragile point prompts. The paper also documents failure modes and hyper-parameter sensitivity, which strengthens its utility as a baseline for future foundation-model adaptation work.

major comments (2)
  1. Table 1 and §4.2: the main claim is reported for N=3, yet several strong baselines (PerSAM, Matcher, IPSeg, T-REX2+SAM) are evaluated under the official 5-shot protocol while PPO is restricted to 1-shot and SAM3 to text-only. The shot-count ablation (Fig. 8) shows saturation after N=3, but a direct head-to-head at identical N (or an explicit statement that 5-shot baselines were re-run at N=3) is needed to make the 8.7-point F_β^w gap fully attributable to the method rather than to unequal reference budgets.
  2. §3.4 and Fig. 7: the operating point τ=0.35 is presented as indispensable for recovering low-contrast camouflage, yet the only sensitivity analysis is a single-dataset sweep. Because GSR is the stage that converts the Top-1 proposal into the final multi-instance mask, a modest domain shift that moves the optimal τ could erase the reported multi-object gains (Table 1, Multi-obj columns). At minimum the paper should quantify how much of the final F_β^w is lost when τ is fixed to the default SAM3 value, or provide a simple adaptive rule.
minor comments (5)
  1. Figure 1 caption refers to “GSR-SAM” while the method is named GFR-SAM throughout; correct the typo.
  2. Eq. (1) and surrounding text contain several spelling/grammar slips (“Subsecquently”, “discared”, “boundding box”); a careful proof-read is needed.
  3. §4.2 states “K=1 in RGCF” while the ablation in Fig. 6 explores K>1; clarify that K=1 is the default used for all main results.
  4. The multi-object subset of R2C7K is small; reporting per-category or bootstrap confidence intervals would strengthen the Multi-obj columns of Table 1.
  5. References [1] and [2] both point to SAM 3; consolidate the citation.

Circularity Check

0 steps flagged

No significant circularity: empirical Generate-Filter-Refine pipeline whose reported metrics are measured against external R2C7K ground truth, not derived by construction from fitted inputs or self-citation lemmas.

full rationale

GFR-SAM is a training-free engineering pipeline (Ice-Seg via modified SAM3 Exemplar Encoder with MSG, RGCF via DINOv3 cosine prototype scoring, GSR via box+text co-prompting) whose central claims are quantitative performance numbers on the public R2C7K split (Table 1: Sm=0.894, Fβw=0.865 at N=3). These numbers are obtained by running the fixed pipeline against held-out pixel masks; no equation reduces a claimed prediction to a parameter that was fitted on the same quantity. Hyper-parameters (τ=0.35, K=1, N=3) are selected by ordinary ablation (Figs. 6–8, Tables 2–4) and do not render the F-measure tautological. Citations to SAM3, DINOv3 and prior Ref-COD work are ordinary related-work references; none supply a uniqueness theorem, ansatz, or load-bearing lemma authored by the present team that forces the result. Failure modes are disclosed rather than definitionally excluded. The derivation chain is therefore self-contained and externally falsifiable.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 3 invented entities

The central performance claim rests on three foundation-model assumptions (SAM3 and DINOv3 features transfer to camouflage without fine-tuning; cosine prototype ranking separates target from distractors; SAM3 refinement with box+text recovers instances) plus three hand-chosen operating points (τ, K, N). The invented modules are engineering constructs, not physical entities; their independent evidence is only the ablations inside this paper.

free parameters (3)
  • confidence threshold τ in GSR = 0.35
    Chosen by sensitivity sweep (Fig. 7) on the evaluation regime; final value 0.35 is required for the reported multi-object recall and overall F_β^w.
  • Top-K candidate retention in RGCF = 1
    Fixed to K=1 after ablation (Fig. 6); larger K trades precision for multi-object recall before refinement.
  • number of reference shots N = 3
    Set to 3 after shot-count ablation (Fig. 8); performance saturates or drops beyond N=3 due to distractor redundancy.
axioms (4)
  • domain assumption SAM3 image and exemplar encoders produce transferable visual prototypes for camouflaged objects without task-specific fine-tuning.
    Invoked throughout §3.2; the entire Ice-Seg stage stands or falls on this transfer.
  • domain assumption DINOv3 patch features yield a cosine similarity heatmap whose region-global contrast ranks true camouflaged instances above background distractors.
    Core of RGCF (§3.3, Eqs. 7–11); Table 3 compares encoders but assumes the ranking property holds for the chosen backbone.
  • ad hoc to paper Discarding E_box and E_pos and applying Mask-guided Spatial Gating yields a position-invariant, background-clean cross-image prompt.
    Explicit design choice in §3.2.2 (Eqs. 5–6); Table 2 shows large drop without MSG, but the construction is paper-specific.
  • ad hoc to paper A lower SAM3 confidence threshold (τ≈0.35) is necessary and sufficient to recover low-contrast camouflaged fragments without catastrophic false positives.
    Stated in §3.4 and Fig. 7; opposite of the usual high-threshold practice for generic objects.
invented entities (3)
  • In-Context Exemplar-guided Segmentation (Ice-Seg) with Mask-guided Spatial Gating no independent evidence
    purpose: Generate candidate masks from an external reference image by rewiring SAM3’s exemplar encoder for cross-image use.
    New module defined in §3.2; independent evidence limited to the paper’s own ablations and qualitative heatmaps.
  • Region-Global Contrastive Filtering (RGCF) no independent evidence
    purpose: Rank and select the single best candidate via DINOv3 prototype-to-patch cosine contrast against global background.
    New scoring module in §3.3; no external validation beyond R2C7K ablations.
  • Geometric-Semantic Refinement (GSR) no independent evidence
    purpose: Recover fine boundaries and multi-instance recall by co-prompting SAM3 with the top-1 box and category text.
    New co-prompting stage in §3.4; Table 4 shows gains over geometric- or semantic-only, still internal evidence.

pith-pipeline@v1.1.0-grok45 · 19329 in / 3440 out tokens · 31929 ms · 2026-07-14T03:33:00.178037+00:00 · methodology

0 comments
read the original abstract

Referring Camouflaged Object Detection (Ref-COD) requires segmenting hidden targets guided by reference cues. While supervised methods are annotation-heavy and training-free approaches via sparse point-prompting are sensitive to localization errors, we propose GFR-SAM, a robust three-stage training-free framework. GFR-SAM shifts the paradigm from fragile point-matching to a "Generate-Filter-Refine" pipeline. First, we introduce In-Context Exemplar-guided Segmentation, empowering SAM3 with cross-image inference to generate candidate masks via holistic visual exemplars, bypassing its native intra-image constraints. Second, a Region-Global Contrastive Filtering module ranks candidates through DINOv3-based prototypical alignment, effectively suppressing background distractors. Finally, a Geometric-Semantic Refinement module synergizes bounding box and text prompts to recover fine-grained boundaries and enhance instance recall. Evaluated on the R2C7K benchmark, GFR-SAM outperforms existing training-free methods by 8.7\% in weighted F-measure ($F_\beta^w$) and competes with supervised state-of-the-art counterparts. Ultimately, this work underscores the potential of unlocking SAM3's latent capability for cross-image In-Context prompting, establishing a robust, training-free paradigm that effectively bridges the gap between general-purpose foundation models and specialized, label-intensive perception tasks without the need for task-specific fine-tuning.

Figures

Figures reproduced from arXiv: 2607.11732 by Jianxin Tian, Liujuan Cao, Shengchuan Zhang, Yilong Yang.

Figure 1
Figure 1. Figure 1: Comparison between the proposed GSR-SAM frame [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overall pipeline of the proposed Generate-Filtering-Refine Framework for Referring Object Segmentation. The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualisation of intermediate and final segmentation of our method. The red box in the candidate mask indicates the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of our GFR-SAM, T-REX2+SAM3 [ [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of similarity heatmaps between the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation on RGCF with different values of [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Ablation on the Number of Shots [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative visualization of failure cases. [PITH_FULL_IMAGE:figures/full_fig_p010_9.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

34 extracted references · 1 canonical work pages

  1. [2]

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, An- drew Huang, Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus Marks, Joseph Greer, Meng Wang, Peize Sun, Roman Rädle, Triantafyllos Afouras, Effrosyni Mavroudi, Katherine Xu, Tsung-Han Wu, Yu Zhou, Lilian...

    Pith/arXiv arXiv 2025

  2. [3]

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. 2021. Emerging properties in self-supervised vision transformers. InProceedings of the IEEE/CVF international conference on computer vision. 9650–9660

    2021

  3. [4]

    Nayee Muddin Khan Dousai and Sven Lončarić. 2022. Detecting humans in search and rescue operations based on ensemble learning.IEEE access10 (2022), 26481–26492

    2022

  4. [5]

    Deng-Ping Fan, Ge-Peng Ji, Ming-Ming Cheng, and Ling Shao. 2021. Concealed object detection.IEEE transactions on pattern analysis and machine intelligence 44, 10 (2021), 6024–6042

    2021

  5. [6]

    Ali Haider, Ghulam Muhammad, Talha Ahmed Khan, Kushsairy Kadir, Mohd Nizam Husen, and Haidawati Mohamad Nasir. 2025. Identification of cam- ouflage military individuals with deep learning approaches DFAN and SINETV2. Scientific Reports15, 1 (2025), 33271

    2025

  6. [7]

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. 2017. Mask r-cnn. InProceedings of the IEEE international conference on computer vision. 2961–2969

    2017

  7. [8]

    Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. 2023. Feature shrinkage pyramid for camouflaged object detection with transformers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5557–5566

    2023

  8. [9]

    Qing Jiang, Feng Li, Zhaoyang Zeng, Tianhe Ren, Shilong Liu, and Lei Zhang

  9. [10]

    InEuropean Conference on Computer Vision

    T-rex2: Towards generic object detection via text-visual prompt synergy. InEuropean Conference on Computer Vision. Springer, 38–57

  10. [11]

    Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. 2023. Segment Anything.arXiv:2304.02643(2023)

    Pith/arXiv arXiv 2023

  11. [12]

    Keshun Liu, Aihua Li, Sen Yang, Changlong Wang, and Yuhua Zhang. 2025. Multi- scale attention and boundary-aware network for military camouflaged object detection using unmanned aerial vehicles.Signal, Image and Video Processing19, 2 (2025), 184

    2025

  12. [13]

    Xuewei Liu, Shaofei Huang, Ruipu Wu, Hengyuan Zhao, Duo Xu, Xiaoming Wei, Jizhong Han, and Si Liu. 2024. Reference prompted model adaptation for referring camouflaged object detection. In2024 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 1–6

    2024

  13. [14]

    Xueyu Liu, Rui Wang, Yexin Lai, Guangze Shi, Feixue Shao, Fang Hao, Jianan Zhang, Jia Shen, Yongfei Wu, and Wen Zheng. 2025. Plug-and-play ppo: An adap- tive point prompt optimizer making sam greater. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 4332–4342

    2025

  14. [15]

    Xuewei Liu, Ziyu Wei, and Jizhong Han. 2025. Leveraging Multimodal Large Language Models for Referring Camouflaged Object Detection. In2025 10th International Conference on Image, Vision and Computing (ICIVC). 196–202. doi:10. 1109/ICIVC66358.2025.11200287

    arXiv 2025

  15. [16]

    Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen

  16. [17]

    Matcher: Segment Anything with One Shot Using All-Purpose Feature Matching.arXiv preprint arXiv:2305.13310(2023)

    Pith/arXiv arXiv 2023

  17. [18]

    Haiyang Mei, Ge-Peng Ji, Ziqi Wei, Xin Yang, Xiaopeng Wei, and Deng-Ping Fan

  18. [19]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Camouflaged Object Segmentation With Distraction Mining. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 8772–8781

  19. [20]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El- Nouby, et al. 2023. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193(2023)

    Pith/arXiv arXiv 2023

  20. [21]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on machine learning. PmLR, 8748–8763

    2021

  21. [23]

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. 2024. SAM 2: Segment Anything in Images and Videos.arXiv preprint arXiv:24...

    Pith/arXiv arXiv 2024

  22. [24]

    Dan Jeric Arcega Rustia, Chien Erh Lin, Jui-Yung Chung, Yi-Ji Zhuang, Ju-Chun Hsu, and Ta-Te Lin. 2020. Application of an image and environmental sensor network for automated greenhouse insect pest monitoring.Journal of Asia-Pacific Entomology23, 1 (2020), 17–28

    2020

  23. [25]

    Kosuke Sakurai, Ryotaro Shimizu, and Masayuki Goto. 2025. Vision and lan- guage reference prompt into sam for few-shot segmentation.arXiv preprint arXiv:2502.00719(2025)

    Pith/arXiv arXiv 2025

  24. [26]

    Oriane Siméoni, Huy V Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Rama- monjisoa, et al. 2025. Dinov3.arXiv preprint arXiv:2508.10104(2025)

    Pith/arXiv arXiv 2025

  25. [27]

    Yanpeng Sun, Jiahui Chen, Shan Zhang, Xinyu Zhang, Qiang Chen, Gang Zhang, Errui Ding, Jingdong Wang, and Zechao Li. 2024. Vrp-sam: Sam with visual reference prompt. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 23565–23574

    2024

  26. [28]

    Lv Tang, Peng-Tao Jiang, Haoke Xiao, and Bo Li. 2025. Towards training-free open-world segmentation via image prompt foundation models.International Journal of Computer Vision133, 1 (2025), 1–15

    2025

  27. [29]

    Liqiong Wang, Jinyu Yang, Yanfu Zhang, Fangyi Wang, and Feng Zheng. 2024. Depth-aware concealed crop detection in dense agricultural scenes. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 17201– 17211

    2024

  28. [30]

    Yu Wen, Shuyong Gao, Shuping Zhang, Miao Huang, Lili Tao, Han Yang, Haozhe Xing, Lihe Zhang, and Boxue Hou. 2025. Referring Camouflaged Object Detec- tion With Multi-Context Overlapped Windows Cross-Attention.arXiv preprint arXiv:2511.13249(2025)

    arXiv 2025

  29. [31]

    Ranwan Wu, Tian-Zhu Xiang, Guo-Sen Xie, Rongrong Gao, Xiangbo Shu, Fang Zhao, and Ling Shao. 2025. Uncertainty-Aware Transformer for Referring Cam- ouflaged Object Detection.IEEE Transactions on Image Processing34 (2025), 5341–5354. doi:10.1109/TIP.2025.3587579

    work page doi:10.1109/tip.2025.3587579 2025

  30. [32]

    Yu-Huan Wu, Zi-Xuan Zhu, Yan Wang, Liangli Zhen, and Deng-Ping Fan. 2025. RefOnce: Distilling References into a Prototype Memory for Referring Camou- flaged Object Detection.arXiv preprint arXiv:2511.20989(2025)

    arXiv 2025

  31. [33]

    Cong Zhang, Kang Wang, Hongbo Bi, Ziqi Liu, and Lina Yang. 2022. Camouflaged object detection via neighbor connection and hierarchical information transfer. Computer Vision and Image Understanding221 (2022), 103450

    2022

  32. [34]

    Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Hao Dong, Peng Gao, and Hongsheng Li. 2023. Personalize Segment Anything Model with One Shot.arXiv preprint arXiv:2305.03048(2023)

    Pith/arXiv arXiv 2023

  33. [35]

    distractor redun- dancy

    Xuying Zhang, Bowen Yin, Zheng Lin, Qibin Hou, Deng-Ping Fan, and Ming- Ming Cheng. 2025. Referring Camouflaged Object Detection.IEEE Transactions on Pattern Analysis and Machine Intelligence47, 5 (2025), 3597–3610. doi:10.1109/ TPAMI.2025.3532440 MM’ 26, October 14-November 1, 2026, Rio de Janeiro, Brazil Yang et al. A Additional Ablation A.1 Ablation on...

    arXiv 2025

  34. [36]

    designing adaptive refinement mechanisms that can dynami- cally adjust the inclusion criteria based on the confidence of the initial segmentation to reduce false positives; and 2) enhancing the robustness of the initial exemplar encoding and matching pro- cess, potentially through multi-scale feature aggregation, to better handle low-contrast and heavily ...