pith. sign in

arxiv: 2606.22546 · v2 · pith:LON2XHAFnew · submitted 2026-06-21 · 💻 cs.CV

Venice-H1: Failure-Aware Query Re-Ranking with Multi-Scale Grid Signatures for Referring Image Segmentation

Nicol\`o Savioli This is my paper

Pith reviewed 2026-06-26 11:20 UTC · model grok-4.3

classification 💻 cs.CV
keywords referring image segmentationquery re-rankingfailure detectionmulti-scale grid signaturesmask selectionpost-hoc moduleDeRIS
0
0 comments X

The pith

Venice-H1 uses multi-scale grid signatures and a failure gate to re-rank candidate masks and close the 3-11% mIoU gap left by argmax selection in referring image segmentation.

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

Modern referring image segmentation models produce several candidate masks per text query yet default to the highest-scoring one, which works on most samples but leaves a persistent error budget on the remaining 7-18 percent. Venice-H1 adds a lightweight post-hoc module that encodes each candidate as compact spatial descriptors pooled on 4x4, 8x8, and 16x16 grids, then passes them through a Transformer re-ranker equipped with a Failure Gate that triggers only when the default choice is likely wrong. On two DeRIS backbones the method raises mIoU on the failure subset by 0.89-1.40 points with positive confidence intervals across every tested split and backbone, while keeping harmful switches below 0.53 percent. The same module transfers zero-shot to medical referring segmentation datasets without any backbone retraining. The added cost is roughly 11.3 million parameters and under one millisecond of latency.

Core claim

The paper claims that encoding each candidate mask through multi-scale grid signatures and routing them to a Transformer re-ranker controlled by a Failure Gate (ROC-AUC 0.78-0.82) enables selective correction of argmax failures, producing consistent mIoU gains on the failure subset, strictly positive 95 percent confidence intervals on all 16 split-backbone combinations, and harmful-switch rates below 0.53 percent, with zero-shot gains observed on medical referring segmentation tasks.

What carries the argument

Multi-scale grid signatures—compact spatial descriptors pooled onto 4x4, 8x8, and 16x16 grids—fed to a Transformer-based re-ranker with an attached Failure Gate that decides whether to override the default argmax selection.

If this is right

  • Failure-case mIoU rises by 1.40 points on DeRIS-L and 0.89 points on DeRIS-B across all evaluated splits and backbones.
  • Harmful-switch rate stays below 0.53 percent while the gate intervenes only on predicted suboptimal cases.
  • Zero-shot transfer produces +1.16 mIoU on MS-CXR and +0.51 mIoU on M3D-RefSeg-2D without any RIS backbone fine-tuning.
  • The module adds about 11.3 million parameters and less than 1 ms of inference latency.

Where Pith is reading between the lines

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

  • The same grid-signature representation could be tested on other tasks that produce multiple spatial outputs, such as referring video segmentation or multi-object tracking.
  • Because the gate is trained only on observed failure patterns, its reliability on entirely new query distributions remains an open measurement.
  • If the signatures prove sufficient to distinguish correct from incorrect masks, they might replace heavier learned re-rankers in resource-constrained settings.

Load-bearing premise

The multi-scale grid signatures together with the trained Failure Gate can detect when the default argmax mask is suboptimal and can generalize that detection to unseen samples without introducing many harmful switches.

What would settle it

A new test set in which the Failure Gate triggers on more than 0.53 percent of samples that were originally correct under argmax, or where the net mIoU change on the failure subset falls to zero or negative with a 95 percent confidence interval that includes zero.

Figures

Figures reproduced from arXiv: 2606.22546 by Nicol\`o Savioli.

Figure 1
Figure 1. Figure 1: Venice-H1 pipeline overview. A frozen DeRIS-L backbone (left, blue) produces N query embeddings qi, mask logit maps Mi, and detection scores si. The feature extraction stage (center) computes mask statistics and multi-scale grid signatures from mask probabilities Pi. The Failure Re-Ranker (right, orange) uses a Transformer-based architecture with two heads: a Failure Gate predicting pˆfail and a Gain Predi… view at source ↗
Figure 2
Figure 2. Figure 2: Multi-scale grid signatures on a RefCOCO example. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Best-query gap vs. actual improvement with 95% [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-sample failure analysis (RefCOCO val). [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: shows the per-split improvement breakdown, [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: IoU distribution of default (Q0) vs. best-query selec￾tions across evaluation splits [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance across all splits. Gray: DeRIS-L default, Blue: DeRIS-L + Venice-H1 (ours), Green hatched: best-query upper bound. Venice-H1 achieves non-negative improvements on all 8 splits [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Coverage–risk trade-off (RefCOCO val). (a) ∆ peaks when gate coverage matches the failure rate (green band). (b) Non-failure regression stays near zero at low cover￾age. 4.5 Qualitative Results [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative re-ranking on RefCOCO val. Each row shows four views: (1) input image with referring expression, (2) ground truth mask (green overlay), (3) default query mask with IoU (red, fails in all cases), and (4) Venice-H1’s corrected selection with IoU and ∆ (blue). In all six examples, the default query produces near-zero IoU while Venice-H1 recovers IoU > 84% [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Ablation study. (a) Multi-scale grids outperform BASE-only and single scale. (b) IoU regression dominates cross-entropy and ListNet. (c) Boundary energy consistently helps [PITH_FULL_IMAGE:figures/full_fig_p009_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Failure gate ROC curves. AUC: 0.78–0.82 across splits. distributions—RefCOCO+ yields macro ∆fail = +1.88 (DeRIS-L) and +1.17 (DeRIS-B) mIoU—confirming that the recovery module is most effective where failures con￾centrate. 9 arXiv:2606.22546 [PITH_FULL_IMAGE:figures/full_fig_p009_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Near-optimal ceiling analysis. (a) DeRIS-L is within 3–5% of the best-query upper bound. (b) Zoomed: Venice-H1 achieves non-negative gains on all splits. (c) 82–93% of samples are already optimal. (d) On failures alone, gains are +0.8–2.2% [PITH_FULL_IMAGE:figures/full_fig_p010_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Cross-domain analysis. Natural images (blue) vs. medical datasets (pink) in a zero-shot setting. (a) Failure rates. (b) Best-query gaps. (c–d) Venice-H1 gains on both domains [PITH_FULL_IMAGE:figures/full_fig_p012_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Unified view across all benchmarks. (a) Full-set ∆ mIoU: natural image splits (blue) and zero-shot medical splits (pink). (b) Failure rate by domain: medical data has ∼10× higher failure rates, creating more opportunities for re-ranking. 12 arXiv:2606.22546 [PITH_FULL_IMAGE:figures/full_fig_p012_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Top-12 failure cases on RefCOCO val. Each cell: default query (left, red overlay) vs. best query (right, green overlay). Ground-truth contour in yellow; IoU gap at top. A correct mask exists among the candidates but is not selected by the default heuristic. 15 arXiv:2606.22546 [PITH_FULL_IMAGE:figures/full_fig_p015_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Extended RefCOCO qualitative gallery (val, 10 additional examples). Input + expression → ground truth (green) → default query (red, with IoU) → Venice-H1 (blue, with IoU and gain). 16 arXiv:2606.22546 [PITH_FULL_IMAGE:figures/full_fig_p016_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Zero-shot medical re-ranking (no fine-tuning). Top: MS-CXR chest X-rays; bottom: M3D-RefSeg-2D 3D medical slices. Default (red), Venice-H1 re-ranked (green), best-query upper bound (blue) [PITH_FULL_IMAGE:figures/full_fig_p017_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Per-sample IoU comparison on medical data. Default (red), Venice-H1 (green), oracle (blue) [PITH_FULL_IMAGE:figures/full_fig_p017_19.png] view at source ↗
read the original abstract

Modern Referring Image Segmentation (RIS) systems generate multiple candidate masks per expression but rely on a simple heuristic--typically the argmax detection score--to select the final output. We identify query selection as a failure-case bottleneck: although heuristic selection succeeds on 82-93% of samples, the residual 7-18% of failures dominate the error budget, leaving a best-query selection gap of 3-11% mIoU. We introduce Venice-H1, a lightweight, backbone-decoupled post-hoc re-ranking module that encodes each candidate through multi-scale grid signatures--compact spatial descriptors pooled onto 4x4, 8x8, and 16x16 grids--and feeds them to a Transformer-based re-ranker with a Failure Gate (ROCAUC 0.78-0.82) that intervenes only when the default choice is likely suboptimal. Instantiated on DeRIS-L and DeRIS-B, Venice-H1 achieves delta_fail of +1.40 and +0.89 mIoU with strictly positive 95% CIs on all 16/16 (split, backbone) pairs and harmful-switch rates below 0.53%. Zero-shot transfer to medical referring segmentation (MS-CXR, M3D-RefSeg-2D) yields +1.16 and +0.51 mIoU without RIS-backbone fine-tuning. The module adds approximately 11.3M parameters and under 1 ms latency.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper introduces Venice-H1, a lightweight backbone-decoupled post-hoc re-ranking module for referring image segmentation (RIS). It encodes multiple candidate masks via multi-scale grid signatures (pooled on 4x4, 8x8, 16x16 grids) and feeds them to a Transformer re-ranker controlled by a Failure Gate (ROCAUC 0.78-0.82) that intervenes only when the default argmax is likely suboptimal. On DeRIS-L and DeRIS-B it reports delta_fail gains of +1.40 and +0.89 mIoU with strictly positive 95% CIs on all 16/16 (split, backbone) pairs, harmful-switch rates below 0.53%, and zero-shot transfer gains of +1.16 and +0.51 mIoU on MS-CXR and M3D-RefSeg-2D without RIS fine-tuning, at a cost of ~11.3M parameters and <1 ms latency.

Significance. If the reported gains and low harmful-switch rates hold under scrutiny, the work offers a practical, low-overhead solution to the query-selection bottleneck that dominates error in current RIS systems (7-18% failure cases). The consistent positive CIs across 16 evaluation settings and the zero-shot medical transfer without backbone retraining would indicate a generalizable failure-detection mechanism that is decoupled from the underlying RIS model.

major comments (1)
  1. [Abstract] Abstract: The central claims of delta_fail gains and harmful-switch rates <0.53% rest on the Failure Gate (ROCAUC 0.78-0.82) generalizing beyond the training distribution of argmax-vs-oracle mismatches. The moderate ROCAUC leaves limited headroom for distribution shift in query phrasing or mask failure modes; without explicit ablations on out-of-distribution failure cases or cross-dataset gate training details, the strictly positive CIs on all 16/16 pairs and the medical zero-shot results could reflect in-distribution behavior rather than robust detection.
minor comments (2)
  1. [Abstract] Abstract: The precise definitions of 'delta_fail' and 'harmful-switch rate' are not stated; a one-sentence definition or pointer to the evaluation protocol would aid immediate comprehension.
  2. [Abstract] Abstract: Grid resolutions are listed as 4x4/8x8/16x16 but the exact pooling operation, feature dimensionality, and how signatures are concatenated before the Transformer are not specified.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the importance of verifying the Failure Gate's generalization. We address this concern directly below, drawing on the existing zero-shot transfer results as primary evidence.

read point-by-point responses

  1. Referee: The central claims of delta_fail gains and harmful-switch rates <0.53% rest on the Failure Gate (ROCAUC 0.78-0.82) generalizing beyond the training distribution of argmax-vs-oracle mismatches. The moderate ROCAUC leaves limited headroom for distribution shift in query phrasing or mask failure modes; without explicit ablations on out-of-distribution failure cases or cross-dataset gate training details, the strictly positive CIs on all 16/16 pairs and the medical zero-shot results could reflect in-distribution behavior rather than robust detection.

    Authors: We acknowledge that an ROCAUC of 0.78-0.82 is moderate and that explicit OOD ablations on failure-case distributions would strengthen the claims. However, the zero-shot transfer experiments on MS-CXR and M3D-RefSeg-2D constitute direct evidence of cross-domain generalization: these datasets use different imaging modalities, query phrasing, and mask failure patterns from the natural-image RIS training distribution, yet the module (including the gate) is applied without any RIS-backbone or re-ranker fine-tuning and still yields +1.16 and +0.51 mIoU gains. The strictly positive 95% CIs across all 16/16 (split, backbone) pairs further indicate that the observed improvements are not artifacts of a single distribution. We will revise the manuscript to (i) explicitly state that the gate was trained only on the source RIS argmax-vs-oracle mismatches and applied zero-shot to the medical sets, and (ii) add a short discussion of why the medical transfer serves as an OOD test for the gate. No new controlled OOD ablation experiments will be added, as the existing cross-domain results already address the core concern. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical results on held-out data

full rationale

The paper introduces a post-hoc re-ranking module (multi-scale grid signatures + Transformer re-ranker + Failure Gate) and reports its performance via direct empirical measurements (mIoU deltas, harmful-switch rates, ROCAUC) on held-out test splits across 16/16 (split, backbone) pairs plus zero-shot medical transfer. No load-bearing derivation, equation, or prediction reduces to its own inputs by construction; the central claims rest on measured generalization rather than self-definition, fitted-input renaming, or self-citation chains. The moderate ROCAUC and generalization assumption are empirical risks, not circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central performance claims rest on the empirical effectiveness of newly introduced grid signatures and the failure gate; these components are design choices whose utility is validated only within the reported experiments.

free parameters (2)
  • Grid resolutions
    4x4, 8x8, 16x16 chosen as the multi-scale pooling sizes for signatures.
  • Failure Gate operating point
    Threshold or decision rule of the gate (ROCAUC 0.78-0.82) is learned from data.
axioms (1)
  • domain assumption Default argmax selection succeeds on only 82-93% of samples, leaving a 7-18% failure set that dominates error.
    Used to motivate the re-ranking module and to define the delta_fail metric.

pith-pipeline@v0.9.1-grok · 5800 in / 1470 out tokens · 37162 ms · 2026-06-26T11:20:19.099613+00:00 · methodology

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

32 extracted references · 22 canonical work pages · 1 internal anchor

  1. [1]

    Vector-based navigation using grid-like representations in artificial agents.Nature, 557:429–433, 2018

    Andrea Banino, Caswell Barry, Benigno Uria, Charles Blundell, Timothy Lillicrap, Dharshan Kumaran, Eleanor Maguire, and Demis Hassabis. Vector-based navigation using grid-like representations in artificial agents.Nature, 557:429–433, 2018. doi:10.1038/s41586-018-0102-6

    work page doi:10.1038/s41586-018-0102-6 2018

  2. [2]

    Navaneeth Bodla, Bharat Singh, Rama Chellappa, and Larry S. Davis. Soft-NMS – improving object detection with one line of code. InICCV, 2017. doi:10.1109/ICCV.2017.593

    work page doi:10.1109/iccv.2017.593 2017

  3. [3]

    Castro, Anton Schwaighofer, Stephanie Hyland, Maria Wetscherek, Tristan Naumann, Aditya Nori, Javier Alvarez-Valle, Hoifung Poon, and Ozan Oktay

    Benedikt Boecking, Naoto Usuyama, Shruthi Bannur, Daniel C. Castro, Anton Schwaighofer, Stephanie Hyland, Maria Wetscherek, Tristan Naumann, Aditya Nori, Javier Alvarez-Valle, Hoifung Poon, and Ozan Oktay. MS-CXR: Making the most of text semantics to improve biomed- ical vision–language processing. InComputer Vision – ECCV 2022, LNCS vol. 13696, pages 1–2...

    2022

  4. [4]

    doi:10.1007/978-3-031-20059-5_1

    work page doi:10.1007/978-3-031-20059-5_1

  5. [5]

    Otaduy, and Dan Casas

    Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. doi:10.1109/CVPR52688.2022.00135

    work page doi:10.1109/cvpr52688.2022.00135 2022

  6. [6]

    DeRIS: Decoupling perception and cognition for enhanced refer- ring image segmentation through loopback synergy.arXiv preprint arXiv:2507.01738, 2025

    Ming Dai, Wenxuan Cheng, Jiang-jiang Liu, Sen Y ang, Wenxiao Cai, Y anpeng Sun, and Wankou Y ang. DeRIS: Decoupling perception and cognition for enhanced refer- ring image segmentation through loopback synergy.arXiv preprint arXiv:2507.01738, 2025. arXiv:2507.01738

    arXiv 2025

  7. [7]

    C 3VG: Multi-task visual grounding with coarse-to-fine consistency constraints

    Ming Dai, Jian Li, Jiedong Zhuang, Xian Zhang, and Wankou Y ang. C 3VG: Multi-task visual grounding with coarse-to-fine consistency constraints. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI),

  8. [8]

    M3d: Advancing 3d medical image analysis with multi-modal large language models.arXiv preprint arXiv:2404.00578, 2024

    Fan Bai, Yuxin Du, Tiejun Huang, Max Q.-H. Meng, and Bo Zhao. M3D: Advancing 3D medical image analysis with multi-modal large language models.arXiv preprint arXiv:2404.00578, 2024. doi:10.48550/arXiv.2404.00578

    work page doi:10.48550/arxiv.2404.00578 2024

  9. [9]

    Otaduy, and Dan Casas

    Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embed- ding for referring image segmentation. InCVPR, 2021. doi:10.1109/CVPR46437.2021.01525

    work page doi:10.1109/cvpr46437.2021.01525 2021

  10. [10]

    Selective classifica- tion for deep neural networks

    Y onatan Geifman and Ran El-Y aniv. Selective classifica- tion for deep neural networks. InNeurIPS, 2017. NeurIPS Proceedings 2017 (Paper 7073)

    2017

  11. [11]

    SelectiveNet: A deep neural network with an integrated reject option

    Y onatan Geifman and Ran El-Y aniv. SelectiveNet: A deep neural network with an integrated reject option. InICML,

  12. [12]

    Francisco Eiras, Kemal Oksuz, Adel Bibi, Philip H. S. Torr, and Puneet K. Dokania. Segment, select, cor- rect: A framework for weakly-supervised referring seg- mentation.arXiv preprint arXiv:2310.13479, 2023. Ac- 13arXiv:2606.22546 OdaxAI Research cepted to ECCV’24 Workshop (Instance-Level Recogni- tion). doi:10.48550/arXiv.2310.13479

    work page doi:10.48550/arxiv.2310.13479 2023

  13. [13]

    Torkel Hafting, Marianne Fyhn, Sturla Molden, May-Britt Moser, and Edvard I. Moser. Microstructure of a spatial map in the entorhinal cortex.Nature, 436:801–806, 2005. doi:10.1038/nature03721

    work page doi:10.1038/nature03721 2005

  14. [15]

    Learning non-maximum suppression

    Jan Hosang, Rodrigo Benenson, and Bernt Schiele. Learning non-maximum suppression. InCVPR, 2017. doi:10.1109/CVPR.2017.701

    work page doi:10.1109/cvpr.2017.701 2017

  15. [16]

    Tailornet: Predict- ing clothing in 3d as a function of human pose, shape and garment style

    Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoqi Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. InCVPR, 2020. doi:10.1109/CVPR42600.2020.01050

    work page doi:10.1109/cvpr42600.2020.01050 2020

  16. [17]

    Mask scoring r-cnn,

    Zhaojin Huang, Lichao Huang, Y ongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring R-CNN. In CVPR, 2019. doi:10.1109/CVPR.2019.00657

    work page doi:10.1109/cvpr.2019.00657 2019

  17. [18]

    Linguistic structure guided context modeling for referring image segmentation

    Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In ECCV, 2020. doi:10.1007/978-3-030-58545-7_4

    work page doi:10.1007/978-3-030-58545-7_4 2020

  18. [19]

    A Piecewise Rotation of the Circle, IPR Maps and Their Connection with Translation Surfaces

    Borui Jiang, Ruixuan Luo, Jiayuan Mao, Tete Xiao, and Yuning Jiang. Acquisition of localization confidence for ac- curate object detection. InECCV, 2018. doi:10.1007/978- 3-030-01264-9_48

    work page doi:10.1007/978- 2018

  19. [20]

    Bokhovkin, S

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Y en Lo, Piotr Dol- lár, and Ross Girshick. Segment anything. InICCV, 2023. doi:10.1109/ICCV51070.2023.00371

    work page doi:10.1109/iccv51070.2023.00371 2023

  20. [21]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. InICCV,

  21. [22]

    doi:10.1109/ICCV.2017.324

    work page doi:10.1109/iccv.2017.324 2017

  22. [23]

    Deepsd: Automatic deep skinning and pose space deformation for 3d garment animation

    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 ICCV, 2021. doi:10.1109/ICCV48922.2021.00986

    work page doi:10.1109/iccv48922.2021.00986 2021

  23. [24]

    Multi-scale representation learning for spatial feature distributions using grid cells (Space2Vec)

    Gengchen Mai, Krzysztof Janowicz, Bo Y an, Rui Zhu, Ling Cai, and Ni Lao. Multi-scale representation learning for spatial feature distributions using grid cells (Space2Vec). InICLR, 2020. OpenReview ICLR 2020

    2020

  24. [25]

    Generation and comprehension of unambiguous object descriptions

    Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. doi:10.1109/CVPR.2016.329

    work page doi:10.1109/cvpr.2016.329 2016

  25. [26]

    Moser, Y asser Roudi, Menno P

    Edvard I. Moser, Y asser Roudi, Menno P . Witter, Clifford Kentros, Tobias Bonhoeffer, and May-Britt Moser. Grid cells and cortical representation.Nature Reviews Neuro- science, 15(7):466–481, July 2014. doi:10.1038/nrn3766. PMID: 24917300

    work page doi:10.1038/nrn3766 2014

  26. [27]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InICML, 2021. PMLR vol. 139

    2021

  27. [28]

    Rylan Schaeffer, Mikail Khona, Tzuhsuan Ma, Cristóbal Eyzaguirre, Sanmi Koyejo, and Ila R. Fiete. Self- supervised learning of representations for space gener- ates multi-modular grid cells. InNeurIPS, 2023. NeurIPS Proceedings 2023

    2023

  28. [29]

    Dataset condensation with distribution matching

    Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhil- iang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mo- hammed, Saksham Singhal, Subhojit Som, and Furu Wei. BEiT-3: Image as a foreign language: BEiT pretraining for vision and vision-language tasks. InCVPR, 2023. doi:10.1109/CVPR56688.2023.01838

    work page doi:10.1109/cvpr56688.2023.01838 2023

  29. [30]

    Otaduy, and Dan Casas

    Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Y andong Guo, Mingming Gong, and Tongliang Liu. CRIS: CLIP- driven referring image segmentation. InCVPR, 2022. doi:10.1109/CVPR52688.2022.02101

    work page doi:10.1109/cvpr52688.2022.02101 2022

  30. [31]

    OneRef: Unified one-tower ex- pression grounding and segmentation with mask referring modeling

    Linhui Xiao, Dunliang Kuang, Siyuan Huang, Shiguang Shan, and Xilin Chen. OneRef: Unified one-tower ex- pression grounding and segmentation with mask referring modeling. InNeurIPS, 2024. NeurIPS Proceedings 2024

    2024

  31. [32]

    Zhao Y ang, Jiaqi Wang, Y ansong Tang, Kai Chen, Heng- shuang Zhao, and Philip H. S. Torr. LAVT: Language- aware vision transformer for referring image segmentation. InCVPR, 2022. doi:10.1109/CVPR52688.2022.01738

    work page doi:10.1109/cvpr52688.2022.01738 2022

  32. [33]

    Venice-H1: Failure-Aware Query Re-Ranking with Multi-Scale Grid Signatures for Referring Image Segmentation

    Licheng Yu, Patrick Poirson, Shan Y ang, Alexander C. Berg, and Tamara L. Berg. Modeling context in referring expressions. InComputer Vision – ECCV 2016, LNCS vol. 9906, pages 69–84. Springer, 2016. doi:10.1007/978-3- 319-46475-6_5. 14arXiv:2606.22546 OdaxAI Research A Extended Qualitative Results Figure 16:Top-12 failure cases on RefCOCO val. Each cell: ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1007/978-3- 2016