Pith. sign in

REVIEW 3 major objections 5 minor 37 references

SCI-CLIP: Segment-Centric Inference with Reference Memory for Training-Free Open-Vocabulary Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Making the segment the unit of every inference stage lets frozen CLIP features segment arbitrary vocabularies without any training.

desk verdict A promising segment-centric inference framework with consistently strong numbers, but the reference-memory provenance is unspecified and the ablation table is messy—fix those before trusting the headline gains. read the letter →

arxiv 2608.05627 v1 pith:EPV3T3Z5 submitted 2026-08-06 cs.CV

Mohamad Zamini , Diksha Shukla This is my paper
classification cs.CV
keywords open-vocabularysemanticsegmentationtraining-freeinferencesegment-centricrepresentationCLIPDINOreferencememoryregion-constrainedaffinitycross-windowKVaggregation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

SCI-CLIP argues that training-free open-vocabulary segmentation fails because it lacks the right unit of inference: patch tokens are organized for global image-text alignment, not for coherent regions. The paper's central claim is that the segment itself should be that unit, governing where features interact, how dense evidence is rebuilt, when context beyond a crop is recovered, and which exemplars are retrieved. Built entirely on frozen CLIP-style features with DINO descriptors and requiring no training, SCI-CLIP reports consistent gains across eight benchmarks, raising VOC21 mIoU from 72.0 to 75.9 and ADE20K from 23.7 to 29.8 under the ViT-B/16 backbone. The thesis a sympathetic reader takes away is that dense open-vocabulary prediction improves not by stronger backbones or stronger retrieval alone, but by aligning every inference stage on one region abstraction.

What carries the argument

The central object is the segment, produced by an auxiliary mask generator (SAM2 by default; Mask2Former, EoMT, and SegFormer also tested). The segment does three jobs in the pipeline. First, an admissibility mask $A \in \{0,1\}^{N\times N}$ zeroes out patch affinities outside a shared region, turning the DINO similarity $S_{ij}=\mathbf{d}_i^\top\mathbf{d}_j$ into scope-constrained attention $\mathbf{W}=\operatorname{softmax}(\tau\hat{S})$ that transports CLIP value tokens (Eqs. 4-9). Second, the same masks pool reconstructed features into reference segment embeddings $\mathbf{e}_r=\tfrac{1}{|r|}\sum_{i\in r}\mathbf{z}_i$ for the offline memory (Eq. 13). Third, they rasterize retrieved segment scores back to pixels for the final fusion $P_{\mathrm{final}}=(\lambda_d P_{\mathrm{dense}}+\lambda_m P_{\mathrm{mem}})/(\lambda_d+\lambda_m)$. Supporting mechanisms are mask merging by DBSCAN on mean region descriptors, a Gaussian locality prior, cross-window KV aggregation that queries a global bank of window keys and values with DINO descriptors, and a three-branch combination $\mathbf{z}_i=\mathbf{z}^{\mathrm{main}}_i+\alpha\mathbf{z}^{\mathrm{spatial}}_i+\beta\mathbf{z}^{\mathrm{semantic}}_i$. The load-bearing identity is that construction, retrieval, and prediction all operate on the same region units.

What would settle it

A direct test of the region-constraint claim: run SCI-CLIP with decoy masks, such as boundaries randomly jittered, regions split into tiny patches, or adjacent objects merged, and compare against intact SAM2 masks; if mIoU stays high while the masks carry no true semantic boundaries, the region-constraint story is wrong. A separate test of the retrieval claim: shuffle the segment-to-label incidence matrix in the reference memory; if the memory-fusion gain survives shuffled label associations, exemplar alignment is not doing the work the paper attributes to it.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that one segment abstraction can carry all stages of dense open-vocabulary inference. SCI-CLIP constrains DINO-based patch affinity with masks from an auxiliary generator to build a region-consistent interaction graph, merges over-fragmented regions, and adds a Gaussian spatial prior; it then reconstructs dense features by propagating value tokens over this graph, selectively recovering long-range evidence across sliding-window crops through cross-window key-value aggregation rather than global attention. The same masks pool the reconstructed features into an offline reference memory of segment embeddings, and retrieved memory scores are fused with the dense prediction at the same segment granularity. Empirically the paper claims this yields spatially coherent, context-aware segmentations that beat prior training-free methods on nearly all of eight benchmarks, with the largest non-retrieval gain from scope reconstruction and the largest late-stage gain from reference-memory fusion.

Load-bearing premise

The load-bearing premise is that the auxiliary mask generator produces region boundaries that match semantic objects: if masks over-fragment or wrongly merge regions, the interaction graph, the reconstructed dense features, and the reference memory all inherit those errors, and the paper itself reports that results shift with the mask generator chosen (Table 3.b).

Editorial extensions

If this is right

  • Training-free OVSS can be improved by reorganizing inference around region structure alone, with no new training data, losses, or backbone updates.
  • Region-constrained interaction is the main structural lever: the stagewise ablation attributes the largest non-retrieval gain to scope reconstruction, so approaches that only sharpen attention or only add retrieval are acting on the wrong layer.
  • Cross-window KV aggregation makes sliding-window inference consistent across crop boundaries, which matters for high-resolution inputs where local reconstruction alone stays crop-dependent.
  • Reference memory aligned with the same segments gives its largest gains on ambiguous benchmarks (V21, C60, Stuff), meaning exemplar correction matters most where dense similarity is least decisive.
  • Because gains persist across mask generators and visual backbones, improvements in either component transfer directly into the pipeline.

Reading between the lines

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

  • A testable transfer the authors do not claim: if the segment abstraction is the real cause, the same scaffolding should improve other frozen-feature dense tasks with the same patch-vs-region mismatch, such as monocular depth or boundary estimation.
  • Their mask-generator comparison shows different region front-ends win on different datasets; a natural extension is per-dataset mask selection or adaptive merging, which could push the reported numbers further without touching the core pipeline.
  • Because memory and prediction share masks, the alignment claim predicts that retrieval gains should erode if reference masks are corrupted or segment-to-label associations are shuffled; that is a cheap way to test whether the memory is genuinely doing exemplar correction.
  • The consistent gains across ViT-B/16, L/14, and H/14 suggest headroom is set by the frozen features, so the reformulation should stack verbatim on any future stronger frozen backbone.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes SCI-CLIP, a training-free, segment-centric inference framework for open-vocabulary semantic segmentation. The method segments an input image into regions, uses those region masks to constrain patch-to-patch affinity (scope reconstruction), adds a Gaussian spatial prior and cross-window key-value aggregation for context recovery, and finally constructs and queries an offline reference memory at the segment level. All stages operate on frozen CLIP-style features without fine-tuning. The authors report consistent improvements over prior training-free methods across eight benchmarks, with the largest gains coming from the reference-memory stage, and provide stagewise and leave-one-out ablations along with sensitivity analyses.

Significance. If the empirical claims hold, SCI-CLIP makes a useful conceptual contribution: it identifies the segment as a unifying inference abstraction for training-free open-vocabulary segmentation, and it bundles feature refinement, context recovery, and exemplar-based correction into one coherent pipeline. The design is simple, requires no training, and the code is promised. The paper also provides ablations that attempt to attribute gains to specific components, which is commendable. The main significance depends critically on whether the reference-memory construction and the per-dataset calibration are free of target leakage and fairly compared with baselines.

major comments (3)
  1. [3.5, 3.6, Eq. (13)-(16)] The provenance of the reference memory is never specified. Eq. (13) pools embeddings from reference images, and Eq. (16) requires a segment-to-label incidence matrix Y_mem, which assumes that every reference segment has a known semantic label. The paper does not state whether these labels come from training-set annotations, validation splits, or the test images themselves, nor whether the reference image set overlaps with the evaluation images. Because Table 4 shows that reference-memory fusion is the single largest contributor to the reported gains (average mIoU from 49.8 to 53.4, ADE20K from 23.3 to 29.8), this omission is load-bearing. The authors must specify the reference image and label source, demonstrate that the test images are excluded, and ideally report the memory-based results with and without this component to bound any leakage effect.
  2. [Table 4, Section 4.4] The stagewise ablation table is internally inconsistent. The row after '+ SR' is labeled '+ GP' twice, with different values (average 49.6 and 49.2), and there is no '+ MM' row even though the text states that mask merging produces a stable improvement. If the second '+ GP' row is in fact the mask-merging row, it shows a decrease in average mIoU and notably on City and V21, contradicting the text. In addition, the text says reference-memory fusion reaches 52.9 average mIoU, while the table reports 53.4. These discrepancies obscure the attribution of gains to individual stages and must be corrected, along with a clear labeling of the rows.
  3. [Table 1.a, Section 4] The method uses dataset-specific calibration of the late-fusion weights λ_d, λ_m and the retrieval scales γ_s, γ_l, with different values for each benchmark. The paper does not describe how these values were selected (e.g., on a validation split) or whether the baseline methods in Table 2 were given analogous per-dataset tuning. Without this information, the comparison in Table 2 may not be apples-to-apples, and the reported gains could partly reflect favorable hyperparameter selection. Please clarify the selection protocol and, if possible, report results for a single fixed configuration or a sensitivity analysis covering the full range.
minor comments (5)
  1. [Section 3.5, Eq. (13)] It is unclear whether the 'reconstructed visual features' used to build the reference memory are the same features produced by the full reconstruction pipeline (Eq. (12)) or the raw frozen CLIP/DINO tokens; please state this explicitly.
  2. [Section 3.4, Eq. (10)] The global keys {k_j} and values {v_j} in the cross-window KV aggregation are not defined in terms of which backbone layer or attention head they come from; please specify the source of these quantities.
  3. [Table 5] The 'Memory' column reports large numbers (e.g., 4585.2) without units or a description of what is measured; please specify the unit (MB, GB) and the measurement protocol.
  4. [Section 4.1, Table 2] The abstract and the opening of Section 4.1 say the method improves across 'eight benchmarks,' but for CLIP B/16 the Cityscapes result (44.7) is lower than the CorrCLIP baseline (45.3); the 'only exception' phrasing is fine, but the abstract's blanket wording should be softened.
  5. [Appendix C] There is a typo in the first paragraph: 'so the it makes the accuracy–resource trade-off explicit' should read 'so that it makes ...'.

Circularity Check

0 steps flagged · score 0.0 of 10

No definitional or self-citation circularity: SCI-CLIP's predictions are computed from frozen features, external masks, and disclosed memory construction; the main open risk is unspecified reference-memory provenance, which is a leakage concern rather than a derivation that reduces to its inputs.

full rationale

Walking the derivation chain, each stage is a genuine function of the image, external masks, and frozen CLIP/DINO features rather than a renamed version of its own output. Dense logits are z_i^T q_k (Eq. 3); reconstructed features are W V-main/W V-spatial plus a mask-conditioned semantic branch (Eqs. 8-12); the reference memory pools these reconstructed features and combines retrieval affinity with a segment-to-label incidence matrix and label-side affinity (Eqs. 13-16). The memory scores are not equal to the dense logits by construction: they require reference segment embeddings, an incidence matrix, and label affinity, so they add information beyond the dense branch. Reusing the same text encoder in both Eq. 3 and Eq. 15 is standard open-vocabulary matching, not a reduction. The paper does disclose per-dataset fusion and retrieval weights (Table 1.a); this is benchmark calibration, not a hidden fit of the headline numbers, and it does not make the derivation circular. There are no load-bearing self-citations: the references are to external methods and datasets, and no uniqueness theorem or prior-work premise is imported from the authors. Two genuine non-circular concerns remain: (i) the provenance of the reference images and labels used to build Y_mem is never specified in Section 3.5, so if evaluation annotations were used the memory gains in Tables 2 and 4 could reflect target leakage; this is a correctness/benchmark-integrity risk, but it is not a claim we can reduce by the paper's own equations to a fitted input. (ii) Table 4 contains a duplicated '+GP' row, omits the '+MM' row promised in the prose, and reports a final average (53.4) inconsistent with the text's 52.9; these are reporting inconsistencies that weaken confidence without demonstrating circularity. The mask-generator dependence is explicitly acknowledged in Section 3.3 and Table 3.b and is an external-input sensitivity, not a self-definitional step.

Assumptions & free parameters 10 free parameters · 4 assumptions · 0 invented entities

The central claim rests on frozen CLIP and DINO embeddings being semantically meaningful at patch level, on the mask generator providing a reliable region decomposition, and on the availability of a labeled reference set. The paper does not introduce any new physical entities. Free parameters include several global and per-dataset hyperparameters, most of which are hand-set or calibrated on the benchmarks.

free parameters (10)
  • Softmax temperature tau for affinity = 4.0 (global)
    Set for all settings; appears in Eq. 8 and controls sharpness of the refined attention.
  • Gaussian prior weight lambda_g = 0.25 (global)
    Set for all settings; appears in Eq. 7 and balances locality against region-constrained affinity.
  • Spatial branch weight alpha = 1.0 (default)
    Chosen by hand after the sensitivity study in Table 1.b; appears in Eq. 12.
  • Semantic branch weight beta = 0.5 (default)
    Chosen by hand after the sensitivity study in Table 1.b; appears in Eq. 12.
  • Dense fusion weight lambda_d = ADE 0.5, City 0.9, Obj/Stf 0.8, C59/60 0.7, V20 0.9, V21 0.8
    Per-dataset calibration values listed in Table 1.a; used in Eq. 17.
  • Memory fusion weight lambda_m = ADE 0.5, City 0.1, Obj/Stf 0.2, C59/60 0.3, V20 0.1, V21 0.2
    Per-dataset calibration values listed in Table 1.a; used in Eq. 17.
  • Segment retrieval scale gamma_s = 60.0 for most datasets, 20.0 for V21
    Listed as Seg. scale in Table 1.a; used in Eq. 14.
  • Label retrieval scale gamma_l = 40.0 or 60.0 depending on dataset
    Listed as Label scale in Table 1.a; used in Eq. 15.
  • Cross-window KV softmax temperature tau_kv = not reported
    Appears in Eq. 10; no value or calibration procedure is given.
  • DBSCAN parameters for mask merging = not reported
    Section 3.3 uses DBSCAN on region features but does not report eps or min_samples.
assumptions (4)
  • domain assumption Frozen CLIP and DINO patch embeddings encode enough semantic structure for dense open-vocabulary prediction.
    The whole method takes these features as given and never updates them; see Section 3.2.
  • domain assumption The auxiliary mask generator produces region proposals whose boundaries correspond to semantic objects.
    Section 3.3 builds the admissibility mask A directly from these regions; Table 3.b shows performance changes with the generator.
  • domain assumption Reference images with known query labels are available and are disjoint from the test set.
    Section 3.5 constructs the memory from reference images but does not state their provenance or label source.
  • domain assumption Standard mIoU benchmark protocols are a valid measure of open-vocabulary segmentation quality.
    The paper evaluates on eight standard benchmarks and reports mIoU without discussing annotation noise or metric limitations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SCI-CLIP: Segment-Centric Inference with Reference Memory for Training-Free Open-Vocabulary Segmentation." pith.science (2026). https://pith.science/paper/EPV3T3Z5

@misc{pith2026260805627,
  author       = {Pith},
  title        = {Pith review of: SCI-CLIP: Segment-Centric Inference with Reference Memory for Training-Free Open-Vocabulary Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EPV3T3Z5}},
  note         = {Machine review of arXiv:2608.05627}
}
read the original abstract

Training-free open-vocabulary segmentation remains limited by a missing inference abstraction. Frozen vision-language features are produced at patch level, yet dense prediction requires a unit that simultaneously governs feature interaction, spatial support, contextual recovery, and retrieval-based correction. We present SCI-CLIP, a segment-centric inference framework built around the principle that the same region abstraction should organize all stages of dense open-vocabulary prediction. SCI-CLIP first induces a region-consistent interaction graph over frozen visual tokens, then reconstructs dense features by propagating values over this graph, augmenting them with selective cross-window support only where local evidence is insufficient. The same segment abstraction is subsequently used to construct and query an offline reference memory, aligning exemplar retrieval with the units on which prediction is made. SCI-CLIP turns frozen CLIP-style features into spatially coherent, context-aware, and retrieval-compatible dense predictions without any training. SCI-CLIP consistently improves the structural quality of dense predictions, the robustness of contextual reasoning, and the alignment of exemplar-based correction, yielding stronger open-vocabulary segmentation across eight benchmarks. Project code is available at: https://github.com/mzamini92/SCICLIP.

Figures

Figures reproduced from arXiv: 2608.05627 by the authors.

Figure 1
Figure 1. Overview of SCI-CLIP. SCI-CLIP consists of an online inference pathway coupled with [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison of SCI-CLIP with existing segmentation methods. More qualitative [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Cross-window KV aggregation restores context across crop boundaries. Each row corre [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Sensitivity of SCI-CLIP to retrieval hyperparameters. We vary the late-fusion reference [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: More qualitative comparison of our method, and the other four methods on City Scapes. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: More qualitative comparison of our method, and the other four methods on COCO Stuff. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: More qualitative comparison of our method, and the other four methods on ADE. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 26 canonical work pages

  1. [1]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Corrclip: Reconstructing patch correlations in clip for open-vocabulary semantic segmentation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  2. [2]

    arXiv preprint arXiv:2603.23030 , year=

    Looking Beyond the Window: Global-Local Aligned CLIP for Training-free Open-Vocabulary Semantic Segmentation , author=. arXiv preprint arXiv:2603.23030 , year=

    work page arXiv

  3. [3]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Training-free open-vocabulary segmentation with offline diffusion-augmented prototype generation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  4. [4]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Emergent open-vocabulary semantic segmentation from off-the-shelf vision-language models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  5. [5]

    European Conference on Computer Vision , pages=

    Proxyclip: Proxy attention improves clip for open-vocabulary segmentation , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    work page 2024

  6. [6]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Grounding everything: Emerging localization properties in vision-language transformers , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  7. [7]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Clip-diy: Clip dense inference yields open-vocabulary semantic segmentation for-free , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    work page

  8. [8]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Image-to-image matching via foundation models: A new perspective for open-vocabulary semantic segmentation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

Show all 37 references
  1. [9]

    Advances in Neural Information Processing Systems , volume=

    Reco: Retrieve and co-segment for zero-shot transfer , author=. Advances in Neural Information Processing Systems , volume=

  2. [10]

    arXiv preprint arXiv:2404.09447 , year=

    knn-clip: Retrieval enables training-free segmentation on continually expanding large vocabularies , author=. arXiv preprint arXiv:2404.09447 , year=

    arXiv

  3. [11]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    DIH-CLIP: Unleashing the Diversity of Multi-Head Self-Attention for Training-Free Open-Vocabulary Semantic Segmentation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  4. [12]

    Proceedings of the Winter Conference on Applications of Computer Vision , pages=

    Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation , author=. Proceedings of the Winter Conference on Applications of Computer Vision , pages=

  5. [13]

    arXiv preprint arXiv:2304.07193 , year=

    Dinov2: Learning robust visual features without supervision , author=. arXiv preprint arXiv:2304.07193 , year=

    arXiv

  6. [14]

    Pattern Analysis, Statistical Modelling and Computational Learning, Tech

    The pascal visual object classes challenge 2012 (voc2012) development kit , author=. Pattern Analysis, Statistical Modelling and Computational Learning, Tech. Rep , volume=

    2012

  7. [15]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    The role of context for object detection and semantic segmentation in the wild , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  8. [16]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Coco-stuff: Thing and stuff classes in context , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  9. [17]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    The cityscapes dataset for semantic urban scene understanding , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  10. [18]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Reproducible scaling laws for contrastive language-image learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  11. [19]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    ResCLIP: Residual Attention for Training-free Dense Vision-language Inference , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  12. [20]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Declip: Decoupled learning for open-vocabulary dense perception , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  13. [21]

    arXiv preprint arXiv:2603.21528 , year=

    PEARL: Geometry Aligns Semantics for Training-Free Open-Vocabulary Semantic Segmentation , author=. arXiv preprint arXiv:2603.21528 , year=

  14. [22]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Feature purification matters: Suppressing outlier propagation for training-free open-vocabulary semantic segmentation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  15. [23]

    European conference on computer vision , pages=

    Sclip: Rethinking self-attention for dense vision-language inference , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  16. [24]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Harnessing vision foundation models for high-performance, training-free open vocabulary segmentation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  17. [25]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Plug-in feedback self-adaptive attention in clip for training-free open-vocabulary segmentation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  18. [26]

    European Conference on Computer Vision , pages=

    Diffusion models for open-vocabulary segmentation , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  19. [27]

    International journal of computer vision , volume=

    Semantic understanding of scenes through the ade20k dataset , author=. International journal of computer vision , volume=. 2019 , publisher=

    2019

  20. [28]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  21. [29]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Emerging properties in self-supervised vision transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  22. [30]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Distilling spectral graph for object-context aware open-vocabulary semantic segmentation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  23. [31]

    European Conference on Computer Vision , pages=

    Clearclip: Decomposing clip representations for dense vision-language inference , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  24. [32]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

    Fossil: Free open-vocabulary semantic segmentation through synthetic references retrieval , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages=

  25. [33]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    ReME: A Data-Centric Framework for Training-Free Open-Vocabulary Segmentation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  26. [34]

    Proceedings of the computer vision and pattern recognition conference , pages=

    Your vit is secretly an image segmentation model , author=. Proceedings of the computer vision and pattern recognition conference , pages=

  27. [35]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Masked-attention mask transformer for universal image segmentation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  28. [36]

    arXiv preprint arXiv:2408.00714 , year=

    Sam 2: Segment anything in images and videos , author=. arXiv preprint arXiv:2408.00714 , year=

    arXiv

  29. [37]

    Advances in neural information processing systems , volume=

    SegFormer: Simple and efficient design for semantic segmentation with transformers , author=. Advances in neural information processing systems , volume=

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.