acceptodds
Under review as a conference paper at ICLR 2027

CANDIDATE-BOX SELECTION MATTERS FOR CROSS- DOMAIN FEW-SHOT OBJECT DETECTION

Abstract

Cross-domain few-shot object detection (CD-FSOD) aims to transfer knowledge from abundant source-domain pretraining to an unseen target domain with only a few labeled samples, while overcoming domain shift and limited target supervision. Grounding DINO has become a mainstream baseline, which selects a fixed number of high-confidence candidate boxes by semantic matching before decoder refinement. In this paper, unlike current works, we identify the selected candidate boxes as an underexplored bottleneck: under domain shift, high-quality boxes may be ranked below the selection budget and discarded, while top-ranked boxes may fail to enclose objects. By comparing in-domain and cross-domain cases, we find that this issue is more severe in cross-domain settings and trace it to vision-language misalignment in candidate selection rather than candidate generation. Motivated by this finding, we propose Support-Conditioned Local Evidence (SCLE), a lightweight and plug-and-play refinement of candidate selection. SCLE constructs foreground and background prototypes from few support samples, builds a local evidence map, and evaluates candidates by in-box purity and inside-outside contrast. A fusion-quota strategy then combines semantically ranked and local-evidence-ranked candidates to better match the target-domain distribution without modifying the detector architecture, adding learnable parameters, or using external data. SCLE can be plugged into different Grounding-DINO-based methods. Extensive experiments validate its effectiveness and demonstrate consistent improvements across state-of-the-art CD-FSOD approaches

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Candidate Selection · Cross-Domain Few-Shot Object Detection · Vision-Language Misalignment · Prototype-Based Adaptation · Transfer Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bycSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates cross-domain few-shot object detection, focusing on the underexplored mechanism of candidate box selection, and highlights that this process may be a critical factor affecting detection performance. The authors find that under domain shift, high-quality candidate boxes may fall outside the selection budget and thus be discarded, while top-ranked boxes might fail to properly enclose the target. By comparing in-domain and cross-domain settings, they observe that this issue is more severe in cross-domain scenarios and attribute it to vision-language misalignment during candidate selection rather than insufficient candidate generation capability. Based on this insight, they propose a lightweight, plug-and-play method called Support-Conditioned Local Evidence (SCLE). SCLE constructs foreground and background prototypes using a small number of support samples, generates local evidence maps, and evaluates candidate boxes based on in-box purity and inside-outside contrast. A fusion-quota strategy then combines candidates ranked by semantic relevance and those ranked by local evidence, enabling better adaptation to the target-domain distribution without modifying the detector architecture, adding learnable parameters, or using external data. The paper presents a relatively novel research angle, and the experimental results validate its effectiveness to some extent. However, the argument regarding the root cause of the problem remains insufficiently substantiated, the methodological novelty is somewhat limited, and the existing experiments are insufficient to fully support the paper's core claims. While the authors primarily attribute candidate selection failures to cross-domain vision-language misalignment, they do not adequately establish this causal relationship. Additionally, the necessity of the three-channel quota strategy and the prototype construction approach, as well as their advantages over simpler alternatives, still require more compelling experimental evidence.

Critical strengths
  1. A Relatively Novel Perspective on the Research Problem This paper identifies pre-decoder candidate-box selection as an underexplored performance bottleneck in cross-domain few-shot object detection (CD-FSOD). By introducing the Coverage Gap metric, the authors quantitatively analyze the discrepancy between candidate-box generation and selection under domain shift. They demonstrate that some high-quality candidates may be discarded due to inaccurate semantic ranking, resulting in additional losses in object coverage. Unlike existing approaches that primarily focus on decoder optimization, fine-tuning strategies, or data augmentation, this work offers a different perspective on understanding performance limitations in CD-FSOD. The identification of this underexplored bottleneck provides meaningful research insights and constitutes an important strength of the paper.
  2. Lightweight Design, Good Compatibility, and Consistent Performance Improvements SCLE modifies only the pre-decoder candidate-box selection mechanism without altering the detector architecture, introducing additional learnable parameters, or requiring external data. This lightweight design offers good compatibility with existing Grounding-DINO-based detection frameworks and demonstrates practical potential. The experiments cover six cross-domain datasets under 1-shot, 5-shot, and 10-shot settings, showing consistent improvements in average mAP on both Grounding DINO and HED. Furthermore, the authors provide additional empirical evidence through candidate-box coverage analysis, component ablation studies, and hyperparameter sensitivity experiments. Although certain limitations remain in the experimental evaluation, the broad experimental coverage and consistent performance gains provide encouraging evidence of the method's practical value.
Critical weaknesses
  1. Insufficient Evidence for the Claimed Vision-Language Misalignment The paper uses the Coverage Gap metric to identify substantial coverage loss during candidate-box selection under domain shift and primarily attributes this issue to vision-language misalignment. However, the experiments do not directly quantify the degree of vision-language misalignment or adequately rule out alternative explanations, such as a weakened correlation between semantic confidence and localization quality, variations in object scale, and background complexity. Therefore, while the results demonstrate a potential bottleneck in candidate-box selection, they do not sufficiently establish vision-language misalignment as its primary cause.
  2. Insufficient Validation of the Three-Channel Quota Mechanism SCLE combines foreground and background prototypes, local evidence scoring, and a three-channel quota mechanism to improve candidate-box selection. However, the paper lacks direct comparisons with simpler alternatives, such as weighted score fusion or straightforward candidate re-ranking. Although Table 5 examines different quota allocations, the evaluation is limited to three datasets under the 1-shot setting. This is insufficient to establish the robustness of the fixed quota strategy across different domains and shot settings. Consequently, the necessity and advantages of the proposed quota mechanism over simpler alternatives remain insufficiently validated.
  3. Potential Confounding Factors in Candidate-Box Coverage Analysis In Section 6.3, the Semantic Top-K and SCLE models are trained independently, potentially resulting in differences in their candidate pools and feature representations. Therefore, the observed coverage improvements may not be solely attributable to the candidate-selection mechanism itself. Although the authors state that differences between the candidate pools are marginal, they do not provide sufficient quantitative evidence to support this assumption. A controlled comparison using the same frozen detector and identical candidate pool would help isolate the contribution of SCLE's selection strategy.
  4. Limited Robustness Evaluation of Support-Set Prototypes SCLE constructs foreground and background prototypes from a small number of labeled target-domain support samples and keeps them fixed during inference. Particularly in the 1-shot setting, these prototypes may be sensitive to support-sample selection, annotation noise, occlusion, and intra-class appearance variations. However, the paper does not systematically evaluate robustness across different support-set samplings or under perturbations to support-sample quality. As a result, the reliability of the proposed approach under challenging real-world conditions remains unclear.
Questions for authors

Question 1: Causal Evidence for Vision-Language Misalignment The paper primarily attributes candidate-box selection failures to vision-language misalignment. However, the Coverage Gap metric only quantifies the loss of candidate coverage during selection and does not directly establish the underlying cause. Could the authors provide additional evidence to support this explanation, such as an analysis of the correlation between semantic matching scores and candidate-box IoU? Question 2: Necessity of the Three-Channel Quota Mechanism Have the authors considered simpler alternatives, such as directly combining semantic scores and local evidence scores through weighted fusion or applying a straightforward candidate-box re-ranking strategy? How do these alternatives compare with the proposed three-channel quota mechanism in terms of detection performance? Additionally, how robust is the fixed quota allocation across different target domains and shot settings? Question 3: Isolating the Contribution of Candidate-Box Selection In Section 6.3, the models using Semantic Top-K and SCLE are trained independently, potentially introducing differences in their candidate pools and feature representations. Could the authors conduct a controlled comparison using the same frozen detector and identical candidate pool to isolate the contribution of the selection mechanism? This would help clarify whether the observed coverage improvements primarily result from the proposed selection strategy rather than differences introduced during training.

2 comments

W1 — Evidence for vision–language misalignment

Thank you for noting that the coverage gap alone does not explain why good candidates are missed. To directly test the effect of vision–language matching on selection, we construct class visual prototypes from target-domain 1-shot supports. We fix the test-image encoder features, all candidate boxes, and the Top-900 budget, mix each original class text feature with its corresponding visual prototype at a 0.75:0.25 ratio, normalize the mixture, temporarily replace the original text feature with the result, and recompute matching scores and selection. We also mix in the negated visual prototype or a random vector with the same weight to test moving away from the prototype and a random perturbation. On UODD, the results are as follows. Retention is measured only for objects that already have an IoU ≥ 0.75 box in the candidate pool:

Original text matchingToward visual prototypeAway from visual prototypeRandom-vector control
Retention of pool-coverable objects84.86%87.81%73.81%85.16%
Net change in retained objects relative to original selection—+59−221+6

With the locations and number of candidate boxes unchanged, moving toward the visual prototype increases retention by 2.95 percentage points, whereas moving away reduces it by 11.04 points; the random control increases it by only 0.30 points. This bidirectional controlled perturbation shows that vision–language matching affects which existing good boxes are selected.

W2 — Value of quota selection across domains and shot settings

To test the suggested score fusion and direct reranking alternative, we rank all candidates by the fused score in Eq. (10) and take the Top-900 without quotas. SCLE instead reserves the 900 slots for semantic, local, and joint evidence. The end-to-end detection results on DeepFish, UODD, Clipart1k, and NEU-DET are as follows (four-domain mean mAP, %):

Selection rule1-shot5-shot10-shot
Semantic score only34.5539.5341.78
Fused score, direct Top-90027.1834.2535.50
SCLE with quotas35.4840.5042.45

Direct fused-score ranking performs below semantic-only selection at every shot setting, whereas SCLE achieves the highest four-domain mean mAP. These results show that direct fused-score ranking does not replace quota selection.

The main results test the fixed allocation across settings. We use the same 450/225/225 quota on all six target domains, at 1/5/10 shots, and with both Grounding DINO and HED, without tuning it by domain or shot. Relative to their respective baselines, six-domain mean mAP improves by 7.4/3.2/3.1 points for Grounding DINO and 1.5/1.0/0.7 points for HED. Table 5 addresses a different question, sensitivity to the quota values: changing the semantic allocation from 300 to 600 changes mAP by at most 1.3 points on the three tested 1-shot datasets.

W3 — Confounding in the coverage analysis

We isolated the selection rule with a frozen-model control. For each test image, we computed the encoder candidates once, then applied semantic Top-900 and SCLE to the same features and boxes under the same 900-box budget. The mean coverage gap across the complete test sets of six domains is shown below (IoU ≥ 0.75; lower is better):

SettingSemantic Top-900SCLEDifference
1-shot0.05300.0419−0.0111
5-shot0.04520.0359−0.0093
10-shot0.04160.0313−0.0103

The six-domain mean improves at every shot setting. On UODD at 1 shot, the gap decreases from 0.108 to 0.071. This matched-pool experiment attributes the measured coverage gain to selection at a fixed representation; Sec. 6.3 reports the related end-to-end comparison.

W4 — Sensitivity to support image quality

We tested support-box annotation error and support-image quality separately.

Box perturbation. For each dataset, we use the 1-shot checkpoint with the highest validation mAP of its paper-selected model (default SCLE for FISH, background expansion 1.8 for UODD, FgOnly for Clipart1k, and BgOnly for NEU-DET). We fix the detector, support images, and clean test set, change each support box center and width/height by at most 5% or 10%, and rebuild the prototypes with three seeds at each perturbation level. This isolates the sensitivity of prototype construction to support-box errors; the detector is not retrained. Detection mAP is:

DatasetPaper result without perturbation5% perturbation10% perturbation
FISH43.342.97 ± 0.0642.93 ± 0.06
UODD23.623.50 ± 0.0023.47 ± 0.06
Clipart1k58.057.77 ± 0.0657.77 ± 0.06
NEU-DET17.016.90 ± 0.0016.90 ± 0.00

The largest numerical difference between the perturbed results and the paper results without perturbation is 0.37 mAP points. We verified that the perturbation changes both the support boxes and rebuilt prototypes.

Support-image perturbation. For each dataset, we use the 1-shot checkpoint with the highest validation mAP of its paper-selected model, fix the detector, support images, and clean test set, add fixed Gaussian pixel noise (standard deviation 0.10 on the [0, 1] scale) only to the support images, rebuild the prototypes from the noisy supports, and evaluate on clean test images; the detector is not retrained:

DatasetPaper result without perturbationNoisy support (no retraining)mAP change
FISH43.343.0−0.3
UODD23.623.4−0.2
Clipart1k58.057.8−0.2
NEU-DET17.016.9−0.1

Under the tested settings, prototypes rebuilt from noisy support images have little effect on the frozen detector; the mAP change is at most 0.3 points across the four domains.

Questions 1–3

Q1. See W1. The two-way matching-score change on fixed candidate boxes directly tests how vision–language matching affects selection.

Q2. See W2. The no-quota fusion comparison tests the value of quotas; the main results across six domains, three shot settings, and two detectors test the fixed allocation across settings.

Q3. See W3. Changing only the selection rule on the same frozen model and candidate pools reduces the six-domain mean coverage gap at all three shot settings.

Reviewer_4qNMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SCLE for cross-domain few-shot object detection. While FT-FSOD improves fine-tuning and decoder design, SCLE focuses on choosing boxes before they enter the decoder. The idea is to consider both how well a box matches the text and how much it looks like the objects in the support set. SCLE builds foreground and background prototypes from a few labeled examples, then uses feature similarity to score each box’s interior and the area around it. It selects 900 boxes in three groups: one based on text scores, one on visual scores, and one on both.

The reported gains on Grounding DINO and HED suggest that better box selection can improve detection.

Critical strengths
  • [S1]. The paper is generally well written and easy to follow. The coverage analysis gives a clear reason to study candidate selection (Table 1).

  • [S2]. The idea of box selection strategy makes sense. Visual features from the support set can help identify useful boxes when text scores are less reliable under domain shift. The proposed approach also adds no trainable parameters.

  • [S3]. The experiments support the method’s effectiveness. SCLE improves average mAP on both Grounding DINO and HED in the 1-, 5-, and 10-shot settings across six domains, although the gains over HED are modest (Table 2).

Critical weaknesses
  • [W1] The overall design is largely heuristic. SCLE combines prototype similarity, a hand-designed score, rank fusion, and fixed quotas (Secs. 5.2–5.5). While these choices make intuitive sense, it is unclear why this particular combination is needed. This makes the entire system quite heuristic.

  • [W2] The comparison with HED needs multiple seeds. The FT-FSOD/HED paper reports results across three official seeds on ODinW-13, with variation of ±0.6 and ±0.7 mAP in the 1- and 3-shot settings. Given SCLE’s modest gains over HED (+1.5/+1.0/+0.7 mAP in Table 2), repeating both methods with the same seeds and support sets, and reporting means and standard deviations, would make the results more convincing.

  • [W3] Key comparisons with simpler selection rules are missing. Tables 3–5 do not isolate the value of ring contrast or quotas. Useful controls would be: (1) using the box-interior score alone, without ring contrast; and (2) selecting the top 900 boxes directly by the combined score in Eq. (10). Table 5’s “no quota” baseline uses only text scores, so it does not test the second case.

  • [W4] Some important details needed for reproducibility are missing. Section 6.1 gives SCLE’s hyperparameters but omits the optimizer, learning-rate schedule, batch size, training length, and checkpoint-selection rule. Please provide the full training configuration and clarify whether HED uses progressive fine-tuning.

Finally, not a weakness, I strongly recommend adding an evaluation on ODinW-13. FT-FSOD/HED already reports results on this benchmark. Comparing HED and HED+SCLE under the same settings and across multiple seeds would provide stronger evidence that SCLE’s gains extend beyond the CD-FSOD domains.

Questions for authors
  • [Q for W2] Is it possible to evaluate both HED and HED+SCLE using the same seeds and support sets, and report means and standard deviations?

  • [Q for W3] Is it possible to add comparisons that isolate the effects of ring contrast and quotas, specifically box-interior scoring alone and Top-900 selection using the combined score in Eq. (10), without quotas?

  • [Q for W4] Is it possible to provide the full training configuration, including the optimizer, learning-rate schedule, batch size, training length, and checkpoint-selection rule, and clarify whether the HED baseline includes progressive fine-tuning?

  • [Q for Suggestion] Is it possible to add an evaluation on ODinW-13, following FT-FSOD’s protocol, to test whether SCLE’s gains extend to more domains?

4 comments

W1 — Why combine these design choices?

We agree that prototype similarity, hand-designed scoring, rank fusion, and fixed quotas are not new components individually. Our contribution is the pre-decoder candidate-selection rule that these components jointly implement. Each component has a role in that rule. Foreground and background prototypes from the support set provide target-domain visual evidence. In-box purity and contrast with the surrounding ring turn token-level similarity into a score for whether a candidate encloses an object. The text score retains candidates favored by the original detector, the local score adds candidates missed by text ranking, and the joint ranking favors candidates supported by both scores. Finally, the 900 slots are divided 450/225/225 across these rankings so that a single fused score cannot displace too many useful semantic candidates.

Simpler selection rules provide a direct comparison. Across four target domains at 1, 5, and 10 shots, selecting only by the local score or directly by the fused score gives lower mean mAP than selecting by the text score alone; the full quota-based SCLE has the highest four-domain mean mAP at every shot setting. W3 gives the separate controls for ring contrast and fusion without quotas, while Table 5 examines quota sensitivity. We revised Secs. 5.2–5.5 to explain this selection logic instead of presenting the components as individually new.

W2 — Multiple seeds

Thank you for this suggestion. We retrained HED and HED+SCLE at 1 shot on Clipart1k, DeepFish, NEU-DET, and UODD using the same support sets, per-domain training configurations, epoch budgets, and checkpoint-selection rule, with seeds 42, 43, and 44. The table reports mean mAP ± standard deviation across the three seeds:

DomainHEDHED+SCLEΔ
Clipart1k56.57 ± 0.3157.70 ± 0.44+1.13
DeepFish42.07 ± 0.8543.10 ± 0.20+1.03
NEU-DET15.17 ± 0.5516.60 ± 0.46+1.43
UODD22.50 ± 0.9623.07 ± 0.06+0.57
Four-domain average34.07 ± 0.4335.12 ± 0.13+1.04

HED+SCLE has higher three-seed mean mAP than HED on all four domains; the four-domain mean rises from 34.07 ± 0.43 to 35.12 ± 0.13. The seed-paired four-domain gains are +0.70, +1.03, and +1.40, positive for all three seeds.

W3 — Comparison with simpler selection rules

Following the reviewer's suggestion, we completed separate ablations for ring contrast and quota selection. Because they use different datasets, we report them separately.

(1) Ring contrast: six-domain mean mAP (%). Removing the ring term leaves only the in-box score Li=PiL_i=P_i; the comparison is full SCLE:

Method1-shot5-shot10-shot
In-box score only, no ring34.744.547.8
Full SCLE35.645.748.6

Full SCLE is higher by 0.9, 1.2, and 0.8 mAP points at the three shot settings, supporting the value of ring contrast for final detection.

(2) Quota selection: four-domain mean mAP (%). Without quotas, the model selects Top-900 directly by the fused score JiJ_i in Eq. (10); the comparison is full SCLE:

Method1-shot5-shot10-shot
Direct fused-score selection, no quota27.1834.2535.50
Full SCLE35.4840.5042.45

Full SCLE is higher by 8.30, 6.25, and 6.95 mAP points. Direct fused-score selection does not replace quota selection in these settings.

W4 — Training details needed for reproduction

For a fair comparison, HED+SCLE follows HED's (FT-FSOD's) training recipe, while Grounding DINO+SCLE follows Grounding DINO's recipe; the full settings are now in Sec. 6.1. HED and HED+SCLE both use progressive two-stage fine-tuning: AdamW (initial learning rate 1×10−41\times10^{-4}, weight decay 0.05), a ReduceOnPlateau schedule, and per-GPU batch size 4. The default training budgets are 16 epochs for FISH, 50 for clipart1k, and 100 for the other domains.

Grounding DINO and Grounding DINO+SCLE use AdamW (initial learning rate 2.5×10−52.5\times10^{-5}, weight decay 1×10−41\times10^{-4}), a 20-iteration warmup followed by MultiStepLR (decays at epochs 19 and 26), per-GPU batch size 1, and 30 epochs, without progressive fine-tuning.

Question on W2: multiple seeds

We retrained HED and HED+SCLE with matched support sets and training settings at 1 shot on four domains, using seeds 42/43/44. Their four-domain mean mAP is 34.07 ± 0.43 for HED and 35.12 ± 0.13 for HED+SCLE; per-domain results are in W2.

Question on W3: simpler rules

We tested the in-box score Li=PiL_i=P_i without ring contrast and Top-900 selection by JiJ_i in Eq. (10) without quotas. W3 reports the retrained mAP comparisons and the dataset scope of each experiment.

Question on W4: training settings

The optimizer, learning-rate schedule, batch size, epoch budget, and checkpoint-selection rule are now in Sec. 6.1 and summarized in W4. Both HED and HED+SCLE use FT-FSOD's progressive two-stage fine-tuning.

ODinW-13 suggestion

Thank you for the suggestion. We evaluated all 13 ODinW-13 datasets under the FT-FSOD 1-shot protocol, using the three officially provided 1-shot support sets. The per-dataset mean mAP across the three splits is as follows:

DatasetHEDHED+SCLE
AerialMaritimeDrone24.526.6
Aquarium45.946.1
CottontailRabbits76.977.5
EgoHands68.369.2
NorthAmericaMushrooms81.879.8
Packages65.067.0
PascalVOC68.769.4
pistols73.273.3
pothole31.032.4
Raccoon68.269.0
ShellfishOpenImages64.067.9
thermalDogsAndPeople76.176.3
VehiclesOpenImages69.270.7
Mean across 13 datasets62.52 ± 0.9663.48 ± 0.68

The overall mean improves by 0.96 mAP points, providing more evidence that SCLE is useful beyond CD-FSOD.

Reviewer_4qNMOpenReview ↗

Follow-up on ODinW-13 Evaluation

Thanks for the clarification and additional experiments. The gain on ODinW-13 1-shot looks maginal. HED reports results for 1, 3, 5, and 10 shots. If I understand correctly, the few-shot training is not very computationally heavy, is it possible to compare both methods across all four settings under the same protocol? This would help show whether the gains over HED are consistent.

Please note that HED also evaluates on RF100-VL. I am not requesting experiments on that benchmark, but HED’s broader evaluation demonstrates its effectiveness.

Thank you for the suggestion. Due to time constraints, we limited our supplementary experiments to the 5-shot setting on all 13 ODinW-13 datasets, using 20 training epochs for both HED and SCLE (100 epochs for HED in the original paper). The per-dataset mAP results are shown below:

DatasetHEDSCLE
AerialMaritimeDrone31.8033.90
Aquarium51.6053.10
CottontailRabbits77.0079.10
EgoHands69.6070.00
NorthAmericaMushrooms87.4088.10
Packages65.9072.50
PascalVOC71.8072.30
pistols72.0072.30
pothole43.2042.00
Raccoon72.1073.40
ShellfishOpenImages62.3067.50
thermalDogsAndPeople78.6082.60
VehiclesOpenImages71.2071.00
Mean across 13 datasets65.7367.52

Across the 13 datasets, SCLE achieves a mean mAP of 67.52, exceeding HED by 1.79 points.

Reviewer_CU3sOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper analyzes the domain-transfer gap associated with box candidate selection and proposes Support-Conditioned Local Evidence (SCLE) to mitigate it. SCLE constructs visual prototypes from support samples to score candidate boxes and combines these scores with a quota-based fusion strategy for candidate selection.

Critical strengths
  • Clear motivation. The paper identifies how Grounding DINO’s default box candidate selection strategy can exacerbate the domain gap in cross-domain settings, particularly through unreliable vision–language alignment.
  • Targeted approach. SCLE introduces a visual prototype-based scoring mechanism designed to mitigate this gap by incorporating evidence from support samples.
  • Supporting empirical results. The reported experiments demonstrate improvements, supporting the value of revisiting candidate selection for cross-domain transfer.
Critical weaknesses
  • Limited methodological novelty. SCLE appears to combine several established components, including prototype construction, foreground/background similarity-based scoring, and relatively simple score-fusion strategies. The paper should more clearly distinguish its novel contributions from existing methods.
  • Unclear significance beyond the specific setting. While I appreciate the motivation, domain-gap analysis, and empirical improvements, the methodological contribution appears incremental. The paper does not sufficiently distinguish SCLE from existing prototype-based scoring and fusion approaches, nor establish the broader significance of the proposed candidate-selection strategy.
Questions for authors
  1. Equation 9: Why is the trade-off coefficient set to 0.5? Please provide a parameter analysis. Could an ablation also establish whether prioritizing P in the ranking is more effective than an additive scoring scheme?
  2. Section 5.5: Could you disentangle the contributions of score fusion and the quota mechanism? In particular, please compare selection using s alone, l alone, and the fused score j, alongside quota-based selection under otherwise matched settings. This would help establish whether the gains come from score fusion or quota-based top-k selection.
  3. Equation 10: How is the fusion weight selected? Is it fixed across datasets and domains, or tuned separately?
  4. Figure 2: What do the labels M, N, and N in the lower-right panel represent? Please clarify their definitions.
2 comments

W1 — Novelty: a change of selection criterion, not the use of prototypes

We agree that prototypes, score fusion, and decoder query selection are not new by themselves. Our paper addresses a specific failure in cross-domain detection: the model generates well-placed candidate boxes, but their text-matching scores are too low for them to enter the decoder. On NEU-DET, the full candidate pool covers 90.0% of objects at IoU ≥ 0.75, while the 900 boxes selected by text-matching score cover only 74.1%; the relative loss on source-domain COCO is just 1.8%. Thus, under domain shift, the selection step discards useful boxes that the model has already generated.

SCLE uses the few labeled support examples to assess whether a candidate encloses an object, together with its text-matching score, before deciding which boxes enter the decoder. In contrast, DE-ViT, CD-ViTO, and LMP mainly use prototypes for class assignment or transfer. A detailed comparison is provided in Table A below:

MethodSignal used for the decisionPipeline stageAdditional target-domain informationImmediate purpose
DE-ViTCosine similarity between query features and class prototypesRegion classification/localization scoring in the open-set detectorClass prototypes from a few support examplesRegion category recognition without detector fine-tuning
CD-ViTOSimilarity to class prototypes / visual promptsClassification and assignment inside the detectorTransferable prompts and detector componentsCategory recognition under domain shift
LMPMultimodal prototype similarityCategory transfer in the detection headMultimodal prototypes and a visual branchCross-domain category transfer
SCLE (ours)Enclosure evidence: interior purity and inside-versus-ring contrast, combined with the semantic scoreFixed Top-K candidate selection before the decoderSupport-derived foreground/background prototypes; no new learnable parametersRetention of candidate boxes likely to enclose targets at high IoU, without making a class decision

W2 — Significance across detector variants, domains, and shot settings

We address the distinction from prior prototype methods in W1 and Table A. The significance of candidate selection is also supported by its performance beyond a single model or dataset: SCLE changes which candidates enter the decoder, adds no learnable parameters, and can be used with Grounding-DINO-based detectors that use different decoding strategies.

Our main results (Table 2) cover six target domains at 1, 5, and 10 shots on both plain Grounding DINO and HED, which uses a parallel decoder. With SCLE, Grounding DINO's six-domain mean mAP rises from 26.3/40.4/43.9 to 33.7/43.6/47.0 at 1/5/10 shots; HED's rises from 34.1/44.7/47.9 to 35.6/45.7/48.6. Figure 3 further shows that SCLE narrows the candidate-coverage gap in most of the 18 domain–shot settings. Together, these results establish pre-decoder selection as a measurable intervention that can complement detector improvements in the CD-FSOD settings tested here. We have made this scope clearer in the abstract and Sec. 1.

Q1 — Eq. (9): the trade-off coefficient and the ranking scheme

In Eq. (9), PiP_i measures evidence inside a candidate box, and CiC_i measures contrast between the box and its surrounding region. To test the factor 0.5, we write the local score as Li=Pi[b+(1−b)Ci]L_i=P_i[b+(1-b)C_i]: bb determines how much interior evidence remains when contrast is weakest. The default b=0.5b=0.5 keeps the contrast term from reducing the interior score by more than half, so high-purity candidates are not discarded when the surrounding evidence is unreliable.

End-to-end validation (HED+SCLE, 1-shot mAP, %). We varied only bb, keeping the other settings as reported for each dataset in the paper:

bFISHUODDclipart1kNEU-DETAvg.
0.343.223.756.715.634.80
0.5 (default)43.323.658.017.035.48
0.743.322.156.916.034.58

Among the three tested values, b=0.5b=0.5 gives the highest four-domain mean mAP (35.48 versus 34.80 and 34.58), validating the default choice. Per-domain results are shown above.

Ranking form (HED+SCLE, 1-shot mAP, %). Under the same training and quota settings, we compared the multiplicative score, the additive score P+CP+C, and ranking by PP first:

Scoring formFISHUODDclipart1kNEU-DETMean
Additive P+C43.223.257.416.034.95
P-first43.323.357.314.934.70
Product (default)43.323.658.017.035.48

The product form uses in-box purity PP as the base and uses ring contrast CC to adjust it: a high ring contrast cannot rescue a candidate with weak in-box evidence, while poor boundary contrast lowers the score of a candidate with strong in-box evidence only moderately. Additive scoring lets CC make up for a low PP, while P-first ranking largely ignores CC; the product keeps both in-box evidence and enclosure information. It gives the highest four-domain mean mAP (35.48), above additive scoring (34.95) and P-first ranking (34.70), so we retain it.

Q2 — Sec. 5.5: separating score fusion from quota selection

We trained separate models for semantic-only selection (s), local-score-only selection (l), fused-score selection without quotas (j), and fused-score selection with quotas (SCLE) on DeepFish (FISH), UODD, clipart1k, and NEU-DET. Their four-domain mean detection mAP is:

Selection rule1-shot5-shot10-shot
Semantic only (s)34.5539.5341.78
Local score only (l)18.6526.5829.85
Fused score, no quota (j)27.1834.2535.50
Fused score + quota (SCLE)35.4840.5042.45

At each shot setting, direct fused-score selection is below semantic-only selection, while SCLE has the highest four-domain mean. So score fusion alone does not improve final detection in these models.

Q3 — Eq. (10): how the fusion weight is chosen

In Eq. (10), ww controls the relative contributions of semantic ranking and local-evidence ranking to the joint ranking. We use a single value, w=0.50w=0.50, for all six target domains and all 1/5/10-shot settings. To examine this choice, we compare w=0.25/0.50/0.75w=0.25/0.50/0.75 on four target domains at 1 shot with seed 42, keeping the other settings fixed and reporting end-to-end detection mAP (%):

Datasetw=0.25w=0.25w=0.50w=0.50w=0.75w=0.75
FISH43.443.343.3
UODD22.923.623.4
clipart1k57.358.057.1
NEU-DET16.617.016.4
Mean35.0535.4835.05

The four-domain mean is highest at w=0.50w=0.50 (35.48), compared with 35.05 at both 0.25 and 0.75. The equal-weight setting is best on UODD, clipart1k, and NEU-DET and is within 0.1 point of the best result on FISH. We therefore use one equal-weight setting rather than selecting a different weight for each domain. These results indicate that SCLE is relatively insensitive to the hyperparameter ww and generalizes well across domains.

Q4 — Figure 2 labels

In the lower-right panel of Fig. 2, MM and the two NN labels show how the 900 candidate slots are allocated: M=αsK=450M=\alpha_sK=450 for candidates ranked by semantic score, N=αlK=225N=\alpha_lK=225 for candidates ranked by local evidence, and N=αjK=225N=\alpha_jK=225 for candidates ranked by the joint score. Hence M+2N=K=900M+2N=K=900; the two NN labels mean that the latter two groups receive 225 slots each. We have clarified this in the revised caption and main text.

Comments on the paper

1 comment

Dear Reviewers,

Thank you for your insightful comments and suggestions. We have carefully addressed the raised concerns in our rebuttal and provided additional clarifications and experimental results where applicable. As the author–reviewer discussion closes on 18 November, we would greatly appreciate it if you could take some time to review our responses; a brief note on whether your concerns have been addressed would help us a great deal.

Thank you again for your time and consideration.

Best regards, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.