acceptodds
Withdrawn from ICLR 2027

PULSE: PERSISTENT UTILIZATION OF LIMITED SUPPORT EVIDENCE FOR CROSS-DOMAIN FEW-SHOT OBJECT DETECTION

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-domain few-shot object detection (CD-FSOD) aims to localize and clas- sify novel objects in unseen domains using only a few annotated target examples. However, fine-tuning-based adaptation can suffer from support underutilization: while limited target examples are used to update detector parameters during fine- tuning, the same target evidence need not remain directly accessible when query predictions are scored. To address this limitation, this work introduces Persis- tent Utilization of Limited Support Evidence (PULSE), a novel framework that continuously leverages annotated target appearances during prediction scoring. Specifically, after target-domain adaptation, PULSE caches multi-view support features in a frozen evidence bank. SSCA is then proposed to align these cached support representations with current query features and measure their support– query agreement. Based on this agreement, SSCA integrates a gated residual correction into the detector’s base class logits, rewarding predictions that are consistent with the observed target support while suppressing confident yet un- supported detections. Importantly, this refinement operates only on classification scores and leaves predicted bounding box coordinates unchanged. Extensive ex- periments across six CD-FSOD benchmarks demonstrate the effectiveness of the proposed approach. PULSE achieves state-of-the-art mean AP under different shot settings. Furthermore, PULSE outperforms the corresponding FT-FSOD results in 16 of the 18 domain–shot settings and matches them in the remaining two, indicating that prediction-time support consultation provides a complementary signal to conventional parameter adaptation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Few-shot

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_aVpZOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose PULSE, a framework designed to address support underutilization in cross-domain few-shot object detection (CD-FSOD) by preserving target evidence beyond fine-tuning. The method operates in two stages: first, fine-tuning a Hybrid Ensemble Decoder (HED) detector and caching multi-view support features in a frozen evidence bank; second, aligning query features with cached supports to apply a gated residual correction to the detector’s class logits via Support-Consistent Score Calibration (SSCA). To refine class scores while protecting the original localization predictions, the authors leverage this support-query agreement to adjust logits without altering predicted boxes. Additionally, the paper evaluates PULSE across six CD-FSOD benchmarks and three shot settings, demonstrating consistent improvements over a strong fine-tuning baseline (FT-FSOD), particularly in overall accuracy and score reliability (ECE, HCFP).

Critical strengths
  1. The proposed PULSE method is simple and easy to implement. The paper calibrates scores by caching support features in Stage I and applying a residual correction to class logits based on support-query agreement in Stage II, without altering the predicted boxes.
  2. The paper is well written and easy to read, with a clear method section. The authors provide a structured problem setup and mathematical formulations, making core mechanisms like support-query matching and residual correction easy to follow.
Critical weaknesses
  1. Limited novelty: PULSE caches support features and performs cosine-based support-query matching, which substantially overlaps with IFC, DE-ViT, and Meta R-CNN. Since the gated residual correction mainly modifies how existing support information is utilized, it would be helpful if the authors could clarify its specific advantages and explicitly highlight how it differentiates from prior works.
  2. Sensitivity to sparse support evidence: PULSE depends on a small and potentially biased support set. Large intra-class variation may give valid queries low support similarity, while visually similar instances from other classes may receive incorrect score enhancement. Maximum-similarity matching can amplify accidental matches, and the frozen evidence bank may drift from evolving query features.
  3. Modest methodological contribution: PULSE modifies only classification logits and leaves bounding boxes unchanged. It cannot correct localization errors, recover missed objects, or improve proposal recall. Its gains mainly come from score re-ranking and false-positive suppression rather than improvements to the detection process itself.
Questions for authors

Please see the weaknesses.

Reviewer_Byj5OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the support underutilization problem in fine-tuning-based cross-domain few-shot object detection (CD-FSOD), where support examples are used during training but are not explicitly referenced during prediction. To address this issue, the paper proposes the Persistent Utilization of Limited Support Evidence (PULSE) model. PULSE first caches multi-view support features into a frozen evidence bank following the target-domain adaptation of the detector. Then, using the Support-Consistent Score Calibration (SSCA) module, the support and query representations are aligned, and the original class logits are updated through gated residual correction based on the agreement between the two representations and the detector confidence. The training process consists of two stages: an initial target-domain adaptation of the detector, followed by the joint training of the detector and the SSCA module. PULSE demonstrates good performance across 1-shot, 5-shot, and 10-shot settings on six different benchmarks.

Critical strengths
  • Rational structural design. To address the support underutilization problem, the authors propose PULSE, a model with a simple yet logical architecture that calibrates existing class logits by leveraging the agreement between the support set and the query set.

  • Consistent performance gains. PULSE shows good performance across 1-shot, 5-shot, and 10-shot settings on six benchmarks and outperforms the baseline (FT-FSOD) in 16 out of the 18 domain–shot combinations, while achieving equivalent performance in the remaining two.

Critical weaknesses
  • Limited novelty and contribution. The use of support during prediction has already been studied in prior works such as support-conditioned detectors and prototype-based approaches. The key distinction of PULSE lies in its residual score refinement design, known as SSCA. However, the paper does not clearly establish whether SSCA offers a meaningful methodological contribution beyond incorporating existing support-utilization ideas into fine-tuning-based CD-FSOD. Even in the problem setup, simply not directly referencing support during prediction does not mean that support information is underutilized. Support information can be incorporated into parameters through fine-tuning, but the paper does not provide sufficient evidence to establish that the lack of explicit support utilization during prediction constitutes a significant performance limitation.

  • Insufficient evidence for SSCA’s effectiveness and unclear experimental settings. In Stage 2, PULSE jointly trains the detector and SSCA for 50 additional epochs. During evaluation, it also applies a support-derived box-size prior and Soft-NMS. However, there is no FT-FSOD baseline with the same additional training and post-processing, nor an ablation isolating the effect of each post-processing operation, making it difficult to determine how much of the performance gain comes from SSCA itself. In the frozen-detector ablation in Table 2, it is also unclear which detector checkpoint was frozen (Stage 1 or Stage 2), whether each SSCA variant was trained separately, and whether the same post-processing was applied to all rows. Furthermore, Base HED in Table 2 achieves a mean AP of 33.08, which differs from FT-FSOD in Table 1 (34.9 AP), whereas Full SSCA matches PULSE in Table 1 on all six domains. Clarification is needed on which checkpoints and training and evaluation conditions are shared between the two experiments.

Questions for authors

Which checkpoint did the frozen HED in Table 2 use, Stage 1 or Stage 2? Also, was each SSCA variant in Table 2 trained separately, or were the variants evaluated by simply removing components from the jointly trained full PULSE model?

Why are the results of FT-FSOD in Table 1 and Base HED in Table 2 different?

What evidence distinguishes the contribution of SSCA itself to the reported performance gains from the effects of the additional 50 training epochs, the box-size prior, and Soft-NMS?

How exactly is the language-score gate (Gn,cG_{n,c}) computed, and what is the rationale behind its design?

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

This submission proposes PULSE for few-shot object detection. It builds on SOTA HED by saving features from labeled support objects and using them to refine detection scores. Specifically, SSCA compares query features with support prototypes and adjusts the original class scores based on their similarity and the detector’s confidence. HED and SSCA are then trained for another 50 epochs while the support bank stays fixed. The full pipeline also uses a support-size prior and Soft-NMS. Experimentally, PULSE reports AP gains of 2.2, 1.0, and 0.6 over FT-FSOD in the 1-, 5-, and 10-shot settings on the CD-FSOD dataset.

Critical strengths
  • [S1] The reported results support the effectiveness of the approach. PULSE improves on FT-FSOD in 16 of the 18 domain–shot settings and matches it in the other two (Table 1). The largest average gain is in the 1-shot setting, where AP improves by 2.2.

  • [S2] Using support examples to refine detection scores is a reasonable idea. The saved support features provide an extra visual reference for assessing the detector’s predictions.

Critical weaknesses
  • [W1] The overall design is largely heuristic, and the scientific insight is missing. PULSE combines feature adapters, prototype pooling, score-correction rules, and Soft-NMS. While the reported gains show that the full system is useful, the submission does not clearly establish why this particular combination is needed or whether its added complexity offers a meaningful advantage over simpler alternatives.

  • [W2] The effects of extra training and post-processing are not isolated. PULSE adds an extra 50 epochs of detector training, a size prior, and Soft-NMS (Secs. 3.4–3.5), while Table 1 uses published FT-FSOD results. The frozen-detector analysis in Table 2 is useful, but does not separately measure these contributions. A HED baseline with the same extra training, plus separate ablations of both post-processing steps, is needed.

  • [W3] The design relies on several hand-designed rules whose necessity is not fully established. For example, four support prototypes, maximum-similarity matching, and threshold-based suppression and gating etc. Table 2 shows stepwise gains. But I think comparisons with simpler choices, such as class-mean prototypes, and sensitivity tests would better justify the added complexity.

  • [W4] The gains need to be checked across multiple runs. Table 1 shows gains of 1.0 and 0.6 AP at 5 and 10 shots, but does not say how much the results varied. FT-FSOD/HED reports standard deviations of 0.6 and 0.7 AP for 1 and 3 shots on ODinW-13. Although this is a different benchmark, but it somehow shows the repeated runs matter. Running for multiple seeds and reporting means and standard deviations, would help show whether the gains hold up, especially at 10 shots, where AP gains remains maginal (47.9 to 48.5).

  • [W5] Some key training details are missing. Appendix A.1 gives the learning rates and training length, but leaves out the optimizer, batch size. It also does not specify the Stage-II loss or which modules are frozen during its first five epochs. These details are needed to reproduce the results.

  • [W6] The presentation needs improvement. For example, in Figure 2, candidate pooling is not clearly connected to the pipeline, and the final “Decoder” is confusing as HED already has logits and boxes. Some key quantities also need clearer definitions: how are the reliability weights in Eq. (5) obtained, and how exactly is the gate GG in Eq. (3) computed? Table 3 gives the HCFP thresholds, but leaves its denominator and matching rule unclear; the ECE binning procedure is also unspecified. These gaps make the method and results harder to follow.

Questions for authors

Please refer to weakness

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

Related papers

Discussion (0)

Sign in to comment.