acceptodds
Under review as a conference paper at ICLR 2027

LoCaR: Localized Canvas Retrieval With Frozen Vision-Language Models

Abstract

Whole-image embeddings can under represent small objects and local attributes that distinguish relevant images for localized queries. We present LoCaR, a training- free framework for localized text-to-image retrieval with frozen vision–language models. LoCaR constructs and utilizes independent proposals (SAM2) focusing on the object so as to get a more localized representation that independently encodes each region using the frozen visual encoder. We further study paired-support adaptation as an optional residual that uses matched text–region exemplars without updating the underlying encoders. We also introduce LocFG, a Visual Genome- derived benchmark designed to evaluate localized retrieval for smaller and more challenging referents. Across several vision language backbones, our results show that improving the region representation itself is the main source of gain.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Zero-shot retrieval, localized image retrieval, text to image retrieval, VLMs, Exemplar Cache

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_unBNOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on localized text-to-image retrieval and proposes LoCaR, a training-free method based on frozen vision-language models. It first uses SAM2 to generate region proposals, and constructs region canvases by cropping the regions and blurring the backgrounds. Each region is independently encoded by a frozen VLM, and the maximum similarity between the text query and the regions is used as the local retrieval score. The score is then fused with the whole-image similarity as the final score. The authors also propose an optional paired-support adaptation method and introduce a new benchmark LocFG. The results on RefCOCOg and LocFG show that the method generally outperforms TextRegion, especially in region retrieval.

Critical strengths
  1. The method is simple and does not require finetuning the VLM. The authors’ evaluations on multiple backbones and datasets show that independently encoding regions can improve retrieval performance.
Critical weaknesses
  1. The novelty of this paper seems limited. The authors mentioned ReCLIP and FGVP in related work, but do not provide direct comparisons with them in the retrieval setting. The proposed method seems quite similar to ReCLIP, for example, region cropping and background blurring, although they focus on different tasks. I think it is important to further explain what is new in LoCaR besides adapting the existing techniques.
  2. The comparison and ablation studies seem limited. TextRegion is mainly compared. However, LoCaR differs from TextRegion in several aspects, including cropping, background blurring, and independent encoding. Also, although the authors compare background blurring with white backgrounds, there is no comparison with plain RGB crops for CLIP and SigLIP2. I think this is important to conduct to tell where the improvements come from, especially Qwen already uses plain RGB crops without background suppression.
Questions for authors
  1. Can the authors provide detailed comparisons between proposed method and previous methods like ReCLIP in the retrieval setting?
  2. Can the authors provide more detailed ablations to show where the improvements come from?
1 comment

Thank you for asking for a direct retrieval comparison and a plain-crop control. We ran both on RefCOCOg. ReCLIP’s two-view scorer gives the strongest region R@1 on both backbones, and plain crops outperform the submitted blurred crops on region R@1. These results clarify what the visual representation contributes and what paired support adds at test time.

W1. Contribution beyond visual prompting. ReCLIP and FGVP established visual prompting techniques that we build on. LoCaR+ adds paired-support adaptation of the retrieval score at test time. A fixed bank stores matched referring expressions and visual regions. For a new query, we measure which support expressions it resembles. For each candidate region, we measure which visual exemplars it resembles. Agreement on the same support pairs adds evidence to the image score. Both encoders stay frozen, and the same residual applies to TextRegion or independently encoded crops.

Table 1 measures the adaptation increment by adding support to the same LoCaR representation. On CLIP, SigLIP2 and Qwen, respectively, support improves region R@1 by 0.67, 0.69 and 0.77 points and region R@10 by 1.77, 1.91 and 1.70 points. Fused R@1 changes by +0.19, −0.06 and +0.15 points. The most consistent benefit of paired support is in the regional score, particularly at R@10.

Q1. Direct comparison with ReCLIP and FGVP. For each of 9,602 expressions, we ranked the same 2,600 gallery images using matched SAM2 proposals, frozen backbones and query embeddings. Each method scores an image by its best proposal. Fused retrieval adds the same whole-image channel within each backbone. We disabled paired support for all methods to compare the visual prompts directly.

ReCLIP combines a tight RGB crop with a full-image view blurred outside the proposal box. We evaluate its isolated-proposal scorer with its spatial-relation rules disabled. FGVP preserves the mask foreground and blurs its surroundings in the full image. Submitted LoCaR applies mask-guided blur inside a padded crop. The table gives R@1 differences in percentage points. Positive values favor the first method.

ComparisonChannelSigLIP2-B/16CLIP-B/16
ReCLIP scorer − LoCaRRegion≈ +10.8≈ +6.7
ReCLIP scorer − LoCaRFused≈ +1.4+0.37
ReCLIP scorer − plain cropRegion+1.72≈ +4.0
LoCaR − FGVPFused≈ +1.8+0.99

On SigLIP2, ReCLIP improves on plain crop for region R@1 (paired McNemar p = 0.000008), while their fused difference is statistically unresolved (p = 0.26). LoCaR exceeds FGVP by about 5.1 region points on SigLIP2. On CLIP, LoCaR’s 0.99-point fused advantage over FGVP is significant (p = 0.0006). The 0.37-point fused gap between ReCLIP and LoCaR on CLIP has no direct paired test.

W2 / Q2. What the ablations isolate. The three SigLIP2 crop variants use identical proposals, 10% padding, 224 px inputs, encoder readout and fusion weight (α = 0.5), with support disabled. Only the treatment of pixels outside the proposal mask changes.

SigLIP2-B/16 representationRegion R@1 (%)Fused R@1 (%)
TextRegion9.5724.69
Plain RGB crop28.3327.61
Mask-guided blur inside crop19.2625.98
White fill outside mask16.5426.29

Plain crops already gain 18.76 region points over TextRegion. Blurring those same crops then costs 9.07 region points and about 1.6 fused points (fused R@1, paired McNemar p ≈ 3 × 10⁻⁷). Table 4 isolates the readout effect. Keeping the blurred canvas fixed and replacing the standard attention-pooled readout with mean patch pooling lowers region R@1 from 19.26% to 10.32%. The crop–TextRegion comparison still changes object scale, visible context and independent encoding together, so it cannot isolate attention locality.

On the proposal-matched 336 px CLIP run, plain crop gives the highest region R@1 of the four representations above, followed by within-crop blur, white fill and TextRegion. Their fused R@1 values span only 0.3 points, with no significant pairwise difference. Qwen uses plain RGB crops throughout, so its support-on versus support-off comparison measures adaptation on that representation. We will revise Section 4.5 to credit the gains already achieved by plain crops and report background suppression as an ablation.

Reviewer_mMuMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes LoCaR, a method for localized text-to-image retrieval that keeps the vision-language encoder frozen. SAM2 generates object proposals for each gallery image, and each proposal is converted into a separate region canvas by preserving the foreground and blurring the background. Each canvas is encoded independently, and its embedding is cached offline. The best query-region similarity is then combined with the whole-image similarity to retrieve relevant images. The paper also introduces an optional paired-support mechanism (LoCaR+) and LOCFG, a new Visual Genome-derived benchmark designed for smaller referents. The problem is interesting and practically relevant, and the improvements with SigLIP2 are notable. We also appreciate that the authors discuss limitations in their comparisons and include a support-set contamination audit, which is useful for researchers working with public RefCOCOg mirrors.

Critical strengths
  1. The paper proposes a new method to encode the background per SAM2 detected concept in the query image.
  2. The results, although shown on one dataset, shows improved performance compared to state-of-the-art TextRegion.
  3. Through thorough ablations, the paper shows the utility of the proposed components.
Critical weaknesses

L1: The contribution of individual components could be isolated more clearly. The idea of independently encoding region canvases is interesting. However, the comparison with TextRegion changes several things at once, including cropping, resizing, background treatment and independent encoding. The authors also acknowledge that their experiments do not isolate the effect of attention locality. Since the paper argues that the region representation is the main source of improvement, it would be helpful to include a few controlled ablations on CLIP and SigLIP2 that separate these factors, for example: ● plain proposal crops without background blur; ● background blur applied to the full image without cropping. This would make it easier to understand which components contribute most to the gains, and would strengthen the explanation for why LoCaR works.

L2: The benchmarking could be strengthened with additional relevant methods. TextRegion is a useful direct baseline, but the evaluation could benefit from additional comparisons with recent methods for small-object or instance-level retrieval. SORCE, which the paper already cites, seems particularly relevant. Could the authors consider evaluating on SORCE-1K or comparing with other suitable retrieval methods where the protocols are compatible? The Qwen results also need to be interpreted carefully. A matched TextRegion baseline is unavailable, and Qwen uses plain proposal crops rather than the background-blurred canvases used with CLIP and SigLIP2. A clearer discussion of these differences would help establish the scope of the claims about performance across backbones.

L3: A few numerical claims and statements should be clarified. Some reported gain ranges do not appear to match the results in Table 1: ● CLIP. The fused LoCaR+ gains over standalone global retrieval are 2.42, 4.90 and 6.15 percentage points at R@1, R@5 and R@10, respectively, rather than the stated range of 2.40-6.05 pp. ● Qwen. The stated range of 9.13-11.69 pp excludes the reported R@5 gain of 12.29 pp. ● Conclusion. The conclusion's range of 2.40-9.13 pp does not match either; even at R@1 alone, the gains are 2.42-9.13 pp. It would help to specify the metric and correct the lower endpoint. The Qwen gains also deserve a note. They are measured against standalone global retrieval (20.17 R@1), but the global channel actually used in fusion reaches 26.48 R@1 (Appendix A.2). Compared with this matched global channel, fused LoCaR+ achieves an improvement of 2.82 pp at R@1. Reporting this matched comparison in the introduction would give readers a clearer picture of the benefit of combining global and regional scores. There is also a mismatch in the statement that paired support improves all three fused metrics for every backbone. For SigLIP2, fused R@1 and R@5 decrease slightly when support is added, although R@10 improves. It would be good to correct these statements and report paired confidence intervals or suitable significance tests for the main comparisons, particularly against TextRegion. This would help readers judge the reliability of the smaller improvements.

L4: The novelty could be positioned more clearly. The canvas construction draws on existing ideas such as proposal cropping and background blurring in ReCLIP and FGVP. The combination may still be useful for gallery-level retrieval, but it would help if the authors more clearly explained what is new beyond adapting these existing ideas to this setting. Related to this, Table 1 shows that unadapted region-only retrieval stays below global retrieval at every rank for all three backbones, so the practical benefit comes from fusing region and global scores. For CLIP, the advantage of the canvas representation over TextRegion largely disappears after fusion. Without paired support, fused TextRegion slightly outperforms LoCaR at R@1 (17.44 vs. 17.38), and with paired support, LoCaR+ leads TextRegion+ by only 0.05 pp (17.57 vs. 17.52). Framing LoCaR explicitly as a complementary channel to global retrieval, and discussing when the canvas representation matters most, would make the contribution more convincing.

Minor Weaknesses: ● Proposal selection and caching: The paper scores only the first 32 proposals per image, although the cache may contain embeddings for additional proposals. Could the authors clarify:

○ how proposals are ordered; ○ why the extra embeddings are stored; ○ whether encoding only the proposals used in the standard setup would reduce preprocessing cost? The additional proposals may be useful for the cap ablation, so clarifying this trade-off would help.

● Hyperparameter and result clarity: The selected settings for K, γ, α and the support IoU threshold could be consolidated into one table for easier reproduction. It would also help to identify the backbone and configuration used for Table 10, and to clarify how its results relate to the corresponding experiments in Table 2.

● Background treatment: In Table 3, white-fill backgrounds give slightly better fused results than blur, although blur gives better standalone region retrieval. This also holds in the comparison without paired support, where both settings use the same fusion weight (e.g., 26.29 vs. 25.99 fused R@1). Since fused retrieval is the main reported metric, could the authors clarify why blur is kept as the default?

Questions for authors
  1. Direct comparison with TextRegion with controlled ablation by changing one component at a time
  2. Comparison with SORCE on SORCE-1k.
  3. Comparison on ILIAS dataset [1] on the text-to-image (t2i) retrieval setting using the reported baselines in ILIAS.
  4. Comparison of compute time (FLOPS or wall-clock timing) against TextRegion.
  5. More emphasis on the novel contribution of LoCAR. As of now, it appears to be encoding the context (background) with every region to bind the region+background/other region context. This however requires repeated feedforward passes (O(N)) for N concepts in the query Image and does not scale well during inference for long queries compared to TextRegion.

[1] https://vrg.fel.cvut.cz/ilias/

2 comments

Thank you for requesting the crop controls and checking Table 1. The controls let us distinguish how a frozen encoder represents a proposal from what paired examples add to its score at test time.

L1 / Q1. Controlled ablations. On RefCOCOg (9,602 queries, 2,600 images), we fixed the SAM2 proposals, text and global embeddings, γ = 0 and α = 0.5. The three SigLIP2 crop variants use identical 10% padding, 224 px inputs and readout.

SigLIP2-B/16Region R@1 (%)Fused R@1 (%)
TextRegion9.5724.69
Plain RGB crop28.3327.61
Mask-guided blur inside crop19.2625.98
White fill outside mask16.5426.29

Plain crop gains 18.76 region points over TextRegion. Blurring that same crop costs 9.07 region points and about 1.6 fused points (paired McNemar p ≈ 3 × 10⁻⁷ for fused R@1). We originally chose blur because it beat white fill on the region channel. White scores slightly higher after fusion (p = 0.32), and plain crop leads both. The matched rerun gives 25.98% for blur versus 25.99% in Table 3. We will revise the rationale for blur in Section 4.5. The crop–TextRegion comparison still changes object scale, context and readout together, so it cannot isolate attention locality.

On the matched 336 px CLIP cache, plain crop also leads within-crop blur on region R@1. The requested full-image blur control preserves the proposal box and blurs its surroundings. It reaches 15.63 region R@1 versus 14.68 for plain crop on CLIP. On SigLIP2, plain crop scores higher. The effect of blur therefore depends on which view is encoded and on the backbone.

L2 / Q2–Q3. Baselines and benchmark scope. ReCLIP’s two-view proposal scorer (spatial rules disabled) leads submitted LoCaR by about 10.8 region R@1 points on SigLIP2 and 6.7 on CLIP. LoCaR leads the FGVP adaptation by about 5.1 region and 1.8 fused points on SigLIP2. Qwen uses plain crops. Its embedding interface lacks the patch-to-text projection our TextRegion implementation requires, so we compare Qwen with its matched global channel.

SORCE-1K’s full-resolution setting is relevant to small-object retrieval. Our completed small-object evaluation is LocFG, where median target area is 3.0% of the image versus 14.5% on RefCOCOg. ILIAS text-to-image retrieval scores multiple photographs of the same physical instance using mAP@1,000 in a distractor gallery. Our source-image recall does not give a comparable score. We have no results on either external benchmark.

L3. Corrected gains. Table 1 fused LoCaR+ gains over standalone global at R@1/5/10 are 2.42/4.90/6.15 points for CLIP, 3.13/4.87/5.94 for SigLIP2, and 9.13/12.29/11.69 for Qwen. The standalone R@1 range is 2.42–9.13. We will lead the introduction with the matched-global R@1 gains of 2.71, 3.13 and 2.82 points, respectively. For Qwen, the progression is 26.48% global → 29.15% with crops → 29.30% with paired support.

L4 / Q5. Contribution of test-time adaptation. LoCaR+ uses a fixed bank of paired expressions and regions. When a query resembles a support expression and a candidate region resembles its paired visual exemplar, their agreement adds evidence to the image score. The same residual applies to TextRegion or crop embeddings with both encoders frozen. At fixed representation, adding support raises region R@1 by 0.67/0.69/0.77 points on CLIP/SigLIP2/Qwen. On SigLIP2, fused R@1/5/10 changes by −0.06/−0.03/+0.16 points. We will correct the all-ranks claim.

The submitted canvas provides a complementary channel to the global score. On CLIP, the 0.05-point fused lead over TextRegion+ is insufficient to establish an advantage. In the matched support-free CLIP run, LoCaR versus TextRegion gives p = 0.57 at fused R@1. On SigLIP2, the new plain-crop region score alone reaches 28.33% against 22.80% global.

Q4 / Q5. Compute and caching. TextRegion pools shared image features. LoCaR generates proposals and encodes each crop offline. A longer query can increase text-encoder cost, but each expression is encoded once and reuses the cached gallery. Both methods score the first 32 proposals in stored SAM2 order (all if fewer). Extra entries permit cap ablations. Encoding only the scored prefix would reduce preprocessing. Gallery caches occupy 78 MB for CLIP and 116 MB for SigLIP2. Matched FLOPs and wall-clock timings are unavailable.

Hyperparameters and Table 10. We will consolidate K, γ, α and the support IoU threshold by configuration. For SigLIP2 blur in Table 3, K = 64, γ = 0.10 and α = 0.5 with support enabled. Table 10 sweeps γ with RefCOCOg anchors and reports region deltas from γ = 0. Its γ = 0.10 row (+0.08/+0.08/+0.12) matches the transferred SigLIP2-B/16 TextRegion+ increments in Table 2, which reports absolute recall. We will identify that correspondence in the caption.

Reviewer_mMuMOpenReview ↗

Reviewer response

I sincerely thank the authors for the results. I sincerely also acknowledge that they presented the results on Plain RGB crop as a strong baseline outperforming their proposed method with and without fusion (Table 3).

Since both the proposed method (blur outside the ROI) and the baseline, Plain RGB crop (crop the ROI) requires the same number of forward passes, and the very simple baseline outperforms the proposed method, I think the requirements for ICLR acceptance are not met.

I stand with my original rating.

Reviewer_A97nOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LoCaR, a training-free method for text-to-image retrieval when the query describes a small object or local detail in an image. For each gallery image, LoCaR uses SAM2 to find object regions, crops each region with the background blurred, and encodes each crop separately with a frozen vision-language model, storing the embeddings in the index offline. At search time, each image is scored by its best-matching region, and this score is combined with the whole-image score; an optional module adds a small correction using matched text-region examples. The authors also build LocFG, a Visual Genome-based benchmark where target objects are smaller than in RefCOCOg. Experiments with CLIP, SigLIP2, and Qwen3-VL-Embedding on RefCOCOg and LocFG show improvements over whole-image retrieval and the TextRegion baseline, and the authors argue that most of the gain comes from the region representation rather than the support module.

Critical strengths
  • The paper introduces LocFG, a benchmark focused on small target objects, with image-disjoint fit/test splits and overlap filtering against RefCOCOg. It could be a useful resource for evaluating localized retrieval in this area.
  • The authors are open about the limitations of their own results. For example, they report a support-set contamination issue in the public RefCOCOg data, the low proposal coverage on LocFG, and the cases where the support module does not help.
Critical weaknesses

Limited novelty...

The core pipeline of LoCaR is already well known in both prior work and practice. It proposes regions in each gallery image, crops each region, encodes each crop with a frozen CLIP-style encoder, stores the region embeddings offline, and ranks images by the best query-region similarity combined with the whole-image score. Also the paper does not cite or compare with the closest existing works.

  • [ref1] studies essentially the same task, retrieving images that contain an object described by an open-vocabulary text query. It starts from the same motivation that a single global embedding misses small objects, ranks each image by the maximum similarity over its local CLIP embeddings, and its supplementary material already uses SAM region proposals to build region-level embeddings. This is very close to both LoCaR and the TextRegion-for-retrieval setting used as the main comparison, but it is neither cited nor compared.
  • [ref2], a blog post from 2022, describes the same indexing idea for real image search systems.
  • [ref3] and [ref4] are public tutorials and code that run SAM on an image, crop each mask as a separate image, embed it with CLIP, store it in a vector database, and search it with a text query. They search within a single image, but extending this to a gallery only changes the scope of the index.

The remaining parts also come from existing work. Cropping with background blur is taken from ReCLIP and FGVP, as the paper itself states.

and also

  • The main claims are stronger than the results.: LoCaR is not compared with TextRegion on Qwen and is not clearly better on CLIP, and the large headline gains (e.g., about 9pp for Qwen) come from a weaker standalone global baseline rather than the matched one used in fusion (under 3pp, Appendix A.2).
  • The source of the gain is not isolated.: Cropping, blur, and independent encoding are changed together without a plain-crop or simple region baseline (e.g., fixed grid crops, [ref1]), even though the white-fill and Qwen plain-crop results suggest blur is not essential.
  • Small differences without significance tests, and unreported cost.: Many conclusions rest on 0.1 to 0.3pp differences with no significance tests, and the indexing cost of up to 32 encoder passes per image plus SAM2 is not reported.

References [ref1] Hila Levi, Guy Heller, Dan Levi, Ethan Fetaya. Object-Centric Open-Vocabulary Image Retrieval with Aggregated Features. BMVC 2023. https://arxiv.org/abs/2309.14999
[ref2] Image search with localization and open-vocabulary reranking using Marqo, yolox, CLIP and OWL-ViT (blog post, Dec 2022). https://dev.to/skelts_tensor_searcher/image-search-with-localization-and-open-vocabulary-reranking-using-marqo-yolox-clip-and-owl-vit-154h
[ref3] Search Within an Image with Segment Anything (LanceDB blog). https://lancedb.com/blog/search-within-an-image-331b54e4285e
[ref4] segment-anything-with-clip (GitHub). https://github.com/curt-park/segment-anything-with-clip

Questions for authors

I would appreciate it if the authors could address the concerns above, especially the novelty issue relative to [ref1] ~ [ref4], in their response.

2 comments

Thank you for suggesting the plain-crop control and pointing us to Levi et al. We will foreground the paper’s adaptation question. Given a frozen region scorer, how much can paired examples improve its predictions at test time?

W1. Novelty relative to [ref1]–[ref4]. Levi et al. [ref1] rank images using local CLIP features and use SAM masks to aggregate dense features. Marqo [ref2] indexes subimages, and [ref3, ref4] search SAM crops with text. We will credit these precedents and the visual prompts of ReCLIP and FGVP. LoCaR+ adds paired-support adaptation at inference. A fixed bank stores labeled expression–region pairs. An image receives extra matching evidence when the query and a candidate region each resemble their respective halves of the same support pairs. We build the bank before testing and keep both encoders frozen. Applying this residual to either TextRegion or independently encoded crops lets us measure adaptation separately from region encoding.

W2. What adaptation contributes. Table 1 gives the following RefCOCOg gains from adding paired support to the same LoCaR representation, in percentage points.

BackboneRegion R@1Region R@10Fused R@1
CLIP-B/16+0.67+1.77+0.19
SigLIP2-B/16+0.69+1.91−0.06
Qwen3-VL-2B+0.77+1.70+0.15

Support improves region recall on all three backbones, while its effect on fused R@1 is small. For Qwen, matched R@1 goes from 26.48% global to 29.15% with crops and 29.30% with paired support. Thus 2.82 points is the total gain and 0.15 is the adaptation increment. The 9.13-point headline used a separate global baseline. Our Qwen interface lacks the patch-to-text projection used by TextRegion, so that comparison covers CLIP and SigLIP2.

W3. Source of the gain. With support disabled, proposals fixed, and fusion weight 0.5, SigLIP2 region/fused R@1 is 9.57/24.69 for TextRegion, 28.33/27.61 for plain crops, 19.26/25.98 for LoCaR blur, and 16.54/26.29 for white fill. Blurring the same crop costs 9.07 region R@1 points. The crop–TextRegion gap also includes changes in object scale and feature readout. ReCLIP’s proposal scorer (spatial rules disabled) leads LoCaR by about 10.8 region points on SigLIP2 and 6.7 on CLIP. LoCaR leads FGVP by about 5.1 on SigLIP2. These results concern matched SAM2 proposals. Direct comparisons with fixed grids and Cluster-CLIP remain open.

W4. Significance and cost. On LocFG, paired support adds 0.36 region R@10 points for Qwen’s selected instruction (image-clustered 95% CI [0.11, 0.61], Appendix A.4). On the matched CLIP cache, the four representations in the crop control span 0.3 fused R@1 points, with no significant pairwise difference. Indexing requires SAM2 and a visual encoding per crop. At query time we score up to 32 cached proposals per image. Table 8 reports gallery/support storage of 78/55 MB for CLIP and 116/80 MB for SigLIP2. We have no matched wall-clock timings.

Q1. What does the paper add beyond region indexing? We quantify what paired examples contribute to a frozen scorer and audit whether those examples are genuinely held out. On 3,000 RefCOCOg test queries, removing 481 shared validation/test rows reduced detected paired overlap from 2.60% to 0.00% at the same threshold (Appendix B). This check matters when the adaptation gains being measured are below one R@1 point.

Reviewer_A97nOpenReview ↗

Thanks

Thank you for the detailed response. The rebuttal reframes the paired-support adaptation as the main contribution, but the submitted paper presents this module as optional and attributes most of the gain to the region representation, which the rebuttal acknowledges has clear precedents in [ref1] to [ref4]. Since neither component can carry the paper's main contribution, I will keep my score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.