acceptodds
Under review as a conference paper at ICLR 2027

From Explanation to Correction: Training-Free Spatial Consistency Intervention for DETR-Based Small Object Detection

Abstract

Small-object detection is fundamentally constrained by the limited and unstable spatial evidence available to a detector. A small object may occupy only a few visual tokens, causing its localization to vary substantially under semantics-preserving transformations or changes in spatial context. Existing methods mainly compensate for this instability through retraining, architectural modification, or black-box prediction aggregation, but rarely exploit the spatial inconsistency already exposed by a frozen detector. We propose Frozen Spatial Consistency Intervention (FSCI), a training- and fine-tuning-free framework that converts observable spatial violations into executable corrections. Counterfactual equivariant explanation identifies localization-sensitive instances through an invertible horizontal transformation and corrects predictions supported by cross-view consistency. Nearest-center ownership slicing recovers local visual detail while assigning each candidate to a unique spatial source, thereby reducing duplicate detections and candidate competition. We evaluate FSCI on five frozen DETR variants—DETR, Deformable DETR, Conditional DETR, DAB-DETR, and Anchor-DETR—across COCO, TinyPerson, and VisDrone2019, forming 15 detector–dataset settings. FSCI improves AP in all 15 settings. Further diagnostic experiments verify the directional and spatial faithfulness of the proposed explanations. These results demonstrate that spatial inconsistency is not merely an inference error, but an actionable source of supervision contained within a frozen detector.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Small Object Detection, DETR, Explainable Artificial Intelligence, Counterfactual Reasoning, Training-Free Inference, Spatial Consistency, Test-Time Intervention

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_PyBbOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Frozen Spatial Consistency Intervention (FSCI). FSCI is a training-free, inference-time framework for frozen DETR-family detectors. Its key idea is that spatial inconsistency under a horizontal flip can serve as a signal for correcting the detector's own small-object predictions. Each image is processed as the original view, its horizontal flip, and four overlapping crops, requiring six forward passes. Flip-view predictions are matched to original predictions by class-constrained Hungarian assignment. Matched boxes are then fused with confidence weights. Crop candidates are assigned to a unique source crop by a nearest-center ownership rule. All candidates are merged through class-aware NMS. FSCI updates no parameters and uses no target-domain annotations. It improves AP in all 15 settings, covering five ResNet-50 DETR variants on COCO val2017, TinyPerson, and VisDrone2019. A mechanistic analysis on COCO includes direction-reversal, random-sign, and budget-matched controls, plus a held-out translation intervention. A boundary analysis shows that instance-level inconsistency does not predict per-instance correction gain.

Critical strengths
  1. Valuable problem and practical setting. Small-object detection with frozen DETR detectors is a real and underexplored problem. The training-free, frozen-detector setting matches practical deployment constraints, such as API-only models, license-restricted weights, and no target-domain annotations.

  2. Consistent empirical gains. FSCI improves AP in all detector–dataset settings, with all hyperparameters shared across settings. The gains concentrate on small objects, which is where the paper's own diagnosis predicts they should appear.

  3. Clear presentation and informative visualization. The paper is well organized and easy to follow. Figure 3 clearly links cross-view attention displacement to correction direction at the instance level. The scale-stratified analyses (Figs. 5–6) make the mechanism easy to follow.

Critical weaknesses
  1. Novelty is not established against the cited techniques. The problem itself is valuable. However, flip augmentation and its benefit for small objects are well documented in prior test-time-augmentation work. The flip-fusion step (Eqs. 3–4) is a two-view special case of Weighted Boxes Fusion, and the slicing step with deduplication is close to SAHI-style inference. The paper cites WBF, SAHI, Soft-NMS, and TTA aggregation as the paradigm it improves upon, yet runs none of them. Without these baselines, the claim that gains come from explanation-guided intervention rather than standard multi-view ensembling cannot be verified.

  2. Six-fold inference cost is neither reported nor justified. Each image requires six forward passes (original, flip, four crops), but no latency, throughput, or memory numbers appear anywhere. No compute-matched alternative is evaluated, such as a single higher-resolution pass or plain flip TTA with standard aggregation. Training-free does not imply cheap inference, but for a frozen-detector wrapper this overhead must be justified.

  3. Baseline detectors are all from 2020–2022. All five are ResNet-50 models predating modern query design. DINO is cited but never evaluated, and RT-DETR is not discussed. The largest gains occur on the weakest baseline, which is also what generic ensembling predicts. Whether the gains persist on modern detectors remains unknown.

  4. Writing details. (a) Lines 192–193: the sentence "A lower ρi or a larger ∥∆bi∥ indicates greater sensitivity to spatial perturbations." is duplicated verbatim. (b) Line 116 (Fig. 2): "Counteractual Image" should be "Counterfactual". (c) Line 456 (Fig. 6e legend): the label "CEF" is not defined in the text; please clarify what it refers to. (d) Line 640: the Appendix placeholder was not removed.

Questions for authors
  1. All main results compare FSCI only to its own single-view baseline. How does FSCI perform against peer test-time methods at matched compute, e.g., flip TTA with WBF, SAHI-style slicing, or a single higher-resolution pass? How much of the gain is attributable to the explanation-guided gating and ownership rule beyond these standard alternatives?

  2. Each image requires six forward passes. What is the latency, throughput, and memory overhead per detector, and how does AP trade off against the number of views? Under what conditions is this preferable to one-time fine-tuning on the public training sets of VisDrone2019 or TinyPerson?

9 comments
AuthorsOpenReview ↗

Response to Question 1

Thank you for requesting these controlled comparisons. We have added Controlled Comparisons with Standard Aggregation and Single-View High-Resolution Control subsections to the appendix, alongside the slicing-based comparison and component analysis. On the complete COCO val2017 set using frozen DETR-R50, we compare methods that reuse the same six source views and retain at most 100 predictions. The controls include class-aware NMS, linear and Gaussian Soft-NMS, and SAHI+flip+WBF with average- and maximum-confidence aggregation. All newly evaluated configurations use an input confidence threshold of 0.01, including a matched-threshold FSCI variant. Gaussian Soft-NMS achieves the highest overall AP among these conventional controls at 39.042, whereas matched-threshold FSCI achieves 42.667 AP, an improvement of 3.625 points. WBF achieves 33.018/35.738 AP with average/maximum confidence. These experiments evaluate WBF under the shared six-view setting rather than flip-only WBF. They match the detector forward-pass budget, source predictions, and output budget, but do not imply identical postprocessing costs. We also evaluate a single high-resolution forward pass whose resolution is selected using training-image latency alone, without validation AP. This control achieves 36.825 AP at 517.06 ms/image, compared with 42.729 AP at 566.96 ms/image for FSCI. Its mean latency is 8.8% lower, which we explicitly acknowledge as an approximate rather than exact runtime match. Regarding attribution, the shared-prediction component analysis provides a more direct comparison: counterfactual correction combined with naive slicing achieves 42.007 AP, while the complete correction and ownership-based candidate-handling procedure achieves 42.729 AP, a difference of 0.722 points. Counterfactual correction alone achieves 42.534 AP, compared with 41.998 AP for the baseline. These differences quantify component contributions under the specified configurations, but are not an additive decomposition of the total gain. In particular, the 0.722-point difference includes the combined ownership and protection procedure and should not be attributed to ownership alone. The direction-reversal, random-intervention, and budget-matched ownership controls further examine whether the prescribed correction direction and spatial assignment matter. Finally, the conventional aggregation controls obtain higher small-object AP but lower medium- and large-object AP than FSCI. Our conclusion is therefore that the proposed procedure provides a better overall accuracy trade-off in these evaluated settings.

AuthorsOpenReview ↗

Response to Question 2

Thank you for raising these practical deployment questions. We have added a Per-Detector Efficiency and View-Count Trade-offs subsection to the appendix, extending the measurements to all seven detectors and reporting the costs of the one-, two-, five-, and six-view DETR configurations. Per-detector inference costs. All detectors are evaluated on the same 200 COCO val2017 images using an NVIDIA GeForce RTX 3060 Laptop GPU, FP32 precision, batch size one, and sequential view processing. After warmup, we perform CUDA-synchronized pipeline measurements with cyclically rotated method order. Timing includes preprocessing, transfers, inference, decoding, and aggregation, but excludes disk access and image decoding. Single-view/FSCI mean latencies are 87.07/517.43 ms for DETR, 277.48/1658.36 ms for Deformable DETR, 104.69/615.07 ms for Conditional DETR, 110.93/660.20 ms for Anchor-DETR, 72.22/426.23 ms for DAB-DETR, 47.24/292.73 ms for DEIMv2-S, and 53.03/330.45 ms for RT-DETRv3-R50. The appendix reports the corresponding throughput and memory measurements. Across these implementations, FSCI increases mean latency by approximately 5.9–6.2×. We explicitly disclose that Deformable DETR uses a local PyTorch attention fallback, so its absolute runtime does not represent the official optimized CUDA implementation. Peak allocated GPU memory remains unchanged within each detector under sequential execution and CPU aggregation. However, we distinguish allocated from reserved memory: Deformable DETR’s peak reserved memory increases from 1930 to 2564 MiB. These measurements exclude CPU memory, CUDA context, and non-PyTorch allocations and do not characterize parallel multiview execution. Accuracy–view-count trade-offs. On complete COCO val2017, the one-view DETR baseline achieves 41.998 AP at 87.07 ms/image. Two-view counterfactual correction achieves 42.534 AP at 174.16 ms/image, five-view ownership slicing achieves 42.407 AP at 432.33 ms/image, and six-view FSCI achieves 42.729 AP at 517.43 ms/image. Thus, two-view correction provides a lower-cost option, whereas the complete method obtains the highest overall and small-object AP among these configurations. We clarify that these are component-defined operating points, not a pure view-count sweep, because the processing rules also differ. When is frozen inference preferable to fine-tuning? We do not claim that FSCI is generally preferable when supervised target-domain adaptation is feasible. TinyPerson and VisDrone2019 provide public training data, making fine-tuning a meaningful alternative. Our evaluation instead addresses the setting where pretrained weights remain frozen and target-domain annotations are not used. FSCI is relevant when parameter updates or a training workflow are unavailable or undesirable, provided additional inference time is acceptable. When labeled data, training resources, and sustained deployment are available, one-time fine-tuning may offer a better amortized accuracy–cost trade-off.

AuthorsOpenReview ↗

Response to Weakness 1

Thank you for this criticism. We agree that horizontal flipping, confidence-weighted box averaging, image slicing, and NMS are established techniques. We do not claim novelty for these individual operations. In particular, the matched-pair fusion in Eqs. (3–4) is a restricted form of confidence-weighted box fusion. Our contribution concerns the combined procedure that establishes class-constrained cross-view correspondence, applies supported localization correction, and regulates crop candidates through spatial ownership, together with controlled analyses of these interventions. As detailed in our response to Question 1, the added aggregation and high-resolution controls examine whether standard alternatives reproduce the observed overall benefit. These results support the complete correction and candidate-handling procedure, rather than establishing weighted averaging as a new operator. We have also tightened the abstract and introduction to align the claims with the evidence. The revised abstract describes cross-view matching as exposing localization discrepancies and supporting correction, rather than implying that inconsistency reliably identifies instances that will benefit. It also acknowledges the limitation of instance-level inconsistency as a predictor of correction gains. In the introduction, we replace the categorical distinction from conventional test-time augmentation with a description of our procedure as building on multiview inference. We further distinguish geometric disagreement and a candidate correction direction from a guarantee that correction will improve an individual prediction. These revisions preserve the empirical contribution while avoiding claims that “explanation” independently causes the gains or that cross-view disagreement reliably predicts per-instance improvement.

Reviewer_FQ2ROpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FSCI, a training-free, inference-time intervention for DETR-based detectors in small-object detection. FSCI consists of two components, Counterfactual Equivariant Explanation and Nearest-Center Ownership Slicing, which augment inference without updating any model weights. Experiments with five frozen DETR variants on three benchmarks validate the effectiveness of both components.

Critical strengths
  1. Convincing empirical results. Under all detector and dataset settings, the proposed method brings improvements with clear margins, which makes the effectiveness of the proposed method well-sound rather than incidental.
  2. Clear presentation. The problem formulation and the overall pipeline are easy to follow, and Figure 2 gives an accurate overview of the method.
  3. Clear report of a negative interaction. Table 4 shows that flip correction with naive slicing leads to only +0.009 AP, compared to the full method's +0.731. Reporting this interaction directly supports the need for ownership-based de-duplication.
Critical weaknesses
  1. Tables 1-3 report only AP, AP50, AP75, and APs, with no APm/APl. Without these metrics, the claim that the full-image branch preserves large objects and that the two branches are complementary are not well supported.
  2. The comparisons in Tables 1-3 are limited to DETR variants published before 2022. Please also include recent SOTA detectors such as DINO, Relation-DETR, MI-DETR, and Mr. DETR, to better demonstrate the method's generalization across architectures.
  3. The paper cites SAHI, WBF, Soft-NMS and better TTA aggregation, but never compares against them. A tuned NMS/Soft-NMS control and SAHI+WBF under matched settings are needed to separate the gain from simply using more views and more candidates.
  4. Some implementation details are missing, e.g., K and the crop/resize policy.
  5. The method requires multiple forward passes per image (original, flip, and crops), yet no latency, memory or throughput figures are reported. Please add a quantitative evaluation of the inference cost.
  6. Please add a controlled comparison against a single pass on an upscaled image at roughly the same cost as FSCI's repeated inference. This would clarify whether the improvement stems from the proposed method rather than from the larger computation.
Questions for authors

See weaknesses above.

6 comments
AuthorsOpenReview ↗

Response to Question 1

Thank you for pointing out that the previously reported metrics were insufficient to support our discussion of large-object preservation and branch complementarity. We have now added two complementary analyses to the appendix: Complete Medium- and Large-Object Results, reporting APₘ and APₗ for all seven detectors across COCO, TinyPerson, and VisDrone2019, and Scale-Specific Component Analysis, examining the individual and combined contributions of the two components using DETR-R50. The complete scale-specific tables supplement the main comparisons using the same frozen weights, inference configurations, and output budgets. They cover all 5,000 COCO validation images, 786 TinyPerson test images, and 548 VisDrone2019 validation images. The results explicitly reveal cross-scale trade-offs: although overall and small-object AP improve across the evaluated settings, medium- and large-object AP do not improve uniformly. For example, DEIMv2-S improves overall and small-object AP on COCO while its medium- and large-object AP decrease. On TinyPerson, medium-object AP improves for all seven detectors, whereas large-object AP exhibits mixed changes. We therefore do not interpret small-object gains as evidence of unconditional performance preservation at other scales. The component analysis provides more direct evidence of branch complementarity. On complete COCO val2017, counterfactual correction improves small-, medium-, and large-object AP from 20.701/45.887/61.059 to 21.485/46.358/61.361, respectively. Ownership slicing produces a larger small-object improvement, reaching 22.419 AP, but slightly reduces medium- and large-object AP. Combining the components achieves 42.729 overall AP and 22.753 small-object AP, exceeding either component alone on both metrics. Compared with ownership slicing alone, the full method also recovers 0.475 medium-object AP and 0.283 large-object AP. All component settings use shared frozen predictions and the same evaluation protocol, without training, fine-tuning, or additional parameter selection. Importantly, the full method’s large-object AP remains 0.110 points below the DETR baseline. The appendix explicitly acknowledges this limitation. The evidence supports the narrower conclusion that full-image correction mitigates the medium- and large-object degradation observed with slicing alone, while the combined procedure improves overall and small-object accuracy. It does not establish guaranteed large-object preservation, uniformly additive component benefits, or the elimination of all cross-scale trade-offs.

AuthorsOpenReview ↗

Response to Question 2

Thank you for this suggestion. We agree that evaluating only earlier DETR variants limits the evidence for applicability to more recent detectors. We have therefore expanded Tables 1–3 to include DEIMv2-S and RT-DETRv3-R50 on COCO, TinyPerson, and VisDrone2019, and updated the corresponding experimental analysis. Both detectors are evaluated with frozen pretrained weights. We apply the same FSCI configuration without training, fine-tuning, or model-specific parameter selection. For a controlled comparison, each baseline and its FSCI counterpart use the same global top-100 prediction budget. On the complete COCO val2017 set, FSCI improves AP from 50.64 to 51.56 for DEIMv2-S and from 52.52 to 53.27 for RT-DETRv3-R50. Their small-object AP increases from 30.89 to 32.30 and from 34.58 to 35.60, respectively. Improvements also extend to the two cross-dataset evaluations. Overall AP increases by 4.66/4.42 points on TinyPerson and 2.11/1.93 points on VisDrone2019 for DEIMv2-S/RT-DETRv3-R50, respectively. These additions provide evidence that the benefits are not restricted to the earlier DETR variants in the original submission. We also retain the reproduction qualifications: RT-DETRv3-R50 uses converted official Paddle inference weights, and its reproduced native baseline is below the published result. All reported improvements are measured against the corresponding reproduced baseline under matched settings.

AuthorsOpenReview ↗

Response to Question 3

Thank you for highlighting the need to distinguish the contribution of FSCI from gains obtained simply by using additional views and candidates. We agree that comparisons with the single-view baseline alone cannot establish this distinction. We have therefore added a Controlled Comparisons with Standard Aggregation subsection to the appendix. The new experiments evaluate class-aware NMS, linear and Gaussian Soft-NMS, and SAHI+flip+WBF with average- and maximum-confidence aggregation on the complete COCO val2017 set of 5,000 images using frozen DETR-R50. All multiview configurations reuse identical source predictions from the original image, its inverse-mapped horizontal reflection, and four overlapping crops. Detector preprocessing, crop geometry, and the final top-100 output budget are matched. Conventional controls apply no FSCI-specific ownership, area filtering, counterfactual correction, or large-box protection. All newly evaluated configurations use an input confidence threshold of 0.01, and we additionally evaluate FSCI with this same threshold. Parameters are fixed without COCO validation-set tuning. Among the evaluated conventional controls, Gaussian Soft-NMS achieves the highest overall AP at 39.042. FSCI with the matched input threshold achieves 42.667 AP, an improvement of 3.625 points. The original FSCI configuration achieves 42.729 AP. WBF achieves 33.018 AP with average-confidence aggregation and 35.738 AP with maximum-confidence aggregation. These results show that the overall benefit persists under matched source views, input filtering, and output budgets. Additional views or the evaluated standard aggregation strategies alone do not reproduce this benefit. Importantly, we also report the cross-scale trade-off rather than claiming uniform superiority. The conventional controls achieve higher small-object AP than FSCI, but substantially lower medium- and large-object AP. FSCI therefore provides a more favorable overall accuracy trade-off, while improving small-object AP over the single-view baseline. The appendix explicitly acknowledges that these results concern the reported fixed configurations, not the best achievable performance of every aggregation method. They support the combined correction and candidate-handling procedure, rather than isolating the contribution of each individual rule.

Reviewer_HBg3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Frozen Spatial Consistency Intervention (FSCI), a training-free inference pipeline for improving small-object detection with frozen DETR-family detectors. It combines detection from 'origin image, horizontal flip, four overlapping image crops' and gives a final prediction through Hungarian matching, nearest-center ownership and additional geometric rules. The paper presents cross-view disagreement as an explanation of localization risk and connects this diagnosis to inference-time correction. Experiments cover five ResNet-50-based DETR variants on COCO, TinyPerson, and VisDrone2019, supplemented by ablations and diagnostic analyses.

Critical strengths
  1. Training free method on frozen detectors. FSCI is a method that does not require training. It identifies useful information by leveraging the characteristic of scale invariance in object detection bounding boxes to filter and optimize the accuracy of the detection boxes.
  2. Significant improvement on the small object detection dataset. The method improves AP in all fifteen detector–dataset combinations. The average gains are approximately 1.01 AP on COCO, 2.39 AP on TinyPerson, and 1.25 AP on VisDrone2019, with an average COCO improvement of approximately 2.57 AP for small objects. The article provides an analysis of current methods and explains their limitations.In particular, it highlights the gap between post-hoc saliency explanations and actionable localization correction, while discussing how generator-based counterfactual approaches can introduce additional computational complexity and entangle semantic changes with geometric effects.
Critical weaknesses
  1. Limited Novelty and Insufficient Comparison with Related Methods The overlapping-crop branch is highly similar to SAHI [1], and the horizontal-flip branch follows standard flip-based TTA. Although the proposed matching and ownership rules may add minor value, the paper lacks a direct comparison with a strong SAHI + flip-TTA baseline under the same six-view setup. Thus the novelty is limited, and the incremental gains of the proposed rules are not well demonstrated. All experiments are conducted on DETR variants proposed between 2020 and 2022, with ResNet-50 as the backbone for every setting. The absence of evaluations on newer, more powerful detectors makes it unclear whether the proposed correction strategy can still yield tangible gains, or if it merely compensates for inherent weaknesses of older models. This undermines the evidence supporting the practical relevance and generalizability of the method.

[1] Fatih Cagatay Akyon, Sinan Onur Altinuc, and Alptekin Temizel. Slicing Aided Hyper Inference and Fine-tuning for Small Object Detection. IEEE International Conference on Image Processing (ICIP), 2022.

[2] Roman Solovyev, Weimin Wang, and Tatiana Gabruseva. Weighted Boxes Fusion: Ensembling Boxes from Different Object Detection Models. Image and Vision Computing, 2021, Article 104117.

  1. Unclear Accuracy–Efficiency Trade-off The approach processes six different views for each input image, which incurs extra detector forward passes compared with the single-view baseline. Nevertheless, the paper reports neither end-to-end inference latency nor throughput. Without such measurements, it is hard to judge whether the accuracy improvements are worthwhile given the added inference overhead, leaving the practical accuracy–efficiency trade-off undetermined.
Questions for authors
  1. For the overlapping-crop and flip TTA design, could you provide results of the SAHI + flip-TTA baseline under the same six-view setting? This will help quantify the actual gains brought by your matching and ownership rules.

  2. Why are all experiments only performed on early DETR variants with ResNet-50 backbones? Can you share results on recent high-performance detectors to verify whether your correction strategy remains effective?

  3. Could you report the end-to-end latency and throughput of your method compared with the single-view baseline? We need these metrics to understand the practical accuracy–efficiency trade-off.

5 comments
AuthorsOpenReview ↗

Response to Question 1

Thank you for highlighting this missing comparison. We agree that horizontal flipping and overlapping crops are established inference operations, and we do not claim either as novel. FSCI contributes a procedure that converts cross-view spatial responses into localization corrections and regulates local candidates through spatial ownership and global–local protection. We have added a subsection entitled “Comparison with Slicing-Based Inference” to the appendix, including a direct comparison on the complete COCO val2017 set of 5,000 images using frozen DETR-R50. SAHI + flip-TTA and FSCI use identical cached predictions from the original image, its horizontal reflection, and four overlapping crops. Detector preprocessing, crop geometry, and the final output budget of 100 detections per image are held constant. We follow the original SAHI benchmark’s released postprocessing configuration: class-wise NMS with an IoU threshold of 0.5 and a confidence threshold of 0.01. As reported in the new comparison table, FSCI achieves 42.729 AP, compared with 37.017 AP for SAHI + flip-TTA, an additional 5.712 points under the same six-view setting. We also explicitly report the trade-off: SAHI + flip-TTA achieves higher small-object AP, at 24.181 versus 22.753, but reduces large-object AP from the original detector’s 61.059 to 50.372. FSCI retains 60.949 large-object AP while improving both overall and small-object AP relative to the original detector. These results show that additional views alone do not explain FSCI’s overall improvement and that FSCI provides a more balanced outcome across object scales.

AuthorsOpenReview ↗

Response to Question 2

Thank you for raising this concern. We agree that evaluating only earlier DETR variants is insufficient to establish applicability to recent detectors. Our original selection used a common ResNet-50 backbone to examine different query and attention mechanisms, but this controlled setting did not address the effect of stronger pretrained representations. We have therefore extended the evaluation to DEIMv2-S and RT-DETRv3-R50 on all three datasets, including the complete COCO val2017 set of 5000 images, TinyPerson test set of 786 images, and VisDrone2019 validation set of 548 images. DEIMv2-S introduces DINOv3-distilled features, extending the evaluation beyond ResNet-based representations. Both detectors use frozen COCO-pretrained weights. We retain their native input resolution and apply the same FSCI parameters across detectors and datasets, without training, fine-tuning, or target-domain calibration. The updated Tables 1–3 report comparisons under the same global output budget of 100 predictions. On COCO, FSCI improves DEIMv2-S from 50.641 to 51.556 AP and RT-DETRv3-R50 from 52.519 to 53.265 AP, with small-object AP gains of 1.416 and 1.024 points, respectively. The improvements are larger under domain shift. Their overall AP gains are 4.658 and 4.424 points on TinyPerson, and 2.111 and 1.935 points on VisDrone2019. These results demonstrate that the observed benefits extend beyond the earlier DETR variants in the original submission. We also checked whether the improvements depend on restricting the original detector’s output budget. Compared with the native top-300 baselines, the overall AP gains remain positive: 0.696 and 0.546 points on COCO, 4.379 and 4.024 points on TinyPerson, and 1.379 and 1.130 points on VisDrone2019, for DEIMv2-S and RT-DETRv3-R50, respectively. Thus, the additional evidence supports modest gains on stronger in-domain baselines and larger gains in the evaluated cross-domain small-object settings. We explicitly disclose a reproducibility limitation for RT-DETRv3: its weights were converted locally from the official Paddle inference checkpoint, and the resulting native top-300 COCO baseline achieves 52.719 AP rather than the published 53.4 AP. We therefore treat its results as qualified evidence for the evaluated converted model. In contrast, the DEIMv2-S native baseline reaches 50.860 AP, closely matching the publicly reported 50.9 AP. We have incorporated these experiments into the main comparison tables and analysis, and updated the experimental statements in the abstract, introduction, and conclusion. The revised evidence supports applicability to the tested recent detectors and representations, while avoiding a claim of universal improvement across all detection architectures.

AuthorsOpenReview ↗

Response to Question 3

Thank you for highlighting the need to quantify the practical accuracy–efficiency trade-off. We agree that training-free inference should not be interpreted as computation-free inference. We have therefore added an Inference Efficiency subsection to the appendix, reporting latency and throughput for single-view DETR-R50, SAHI+flip-TTA, and FSCI. All methods are measured on the same NVIDIA GeForce RTX 3060 Laptop GPU using FP32 precision and a batch size of one. We randomly sample 200 COCO val2017 images and repeat the measurements three times, obtaining 600 timed runs per method after warmup. Method order is rotated across images and repetitions, with CUDA synchronization before and after each timed pipeline. Measurements include view construction, preprocessing, CPU–GPU transfers, model inference, decoding, coordinate mapping, and postprocessing. Disk access and image decoding are excluded by preloading images into memory. Throughput is measured under sequential inference. Single-view DETR-R50 requires 102.09 ms/image and processes 9.80 images/s, whereas FSCI requires 607.13 ms/image and processes 1.65 images/s. Thus, FSCI incurs approximately 5.95× the baseline latency. As reported in the full COCO val2017 evaluation, this additional computation improves AP from 42.00 to 42.73 and small-object AP from 20.70 to 22.75. Under the same six-view budget, SAHI+flip-TTA requires 615.55 ms/image, indicating that FSCI has comparable total runtime. The additional CPU fusion stage of FSCI averages 3.93 ms, approximately 0.65% of its total latency. Most runtime is therefore spent outside this final fusion stage, including additional view processing and repeated detector inference. These measurements clarify both the benefit and the limitation of our approach. FSCI removes the need for retraining, but it introduces substantial inference overhead and is not a real-time acceleration method in the current implementation. Its intended use is in settings where additional inference computation is acceptable in exchange for improved detection accuracy without updating model parameters. We now explicitly discuss this trade-off in the appendix rather than implying that training-free correction has negligible deployment cost.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Summary of Contributions and Discussion

Dear Area Chair,

We thank you and the reviewers for the constructive evaluation and discussion. We summarize below the central contribution of our work and the evidence strengthened during the review process.

  1. Connecting spatial diagnosis to actionable correction. FSCI studies how spatial inconsistencies in frozen detectors can be converted into useful inference-time interventions for small-object detection. Its contribution is not horizontal flipping, slicing, or box averaging individually. Rather, it combines geometrically aligned instance correction with explicit crop-source ownership, and evaluates why these interventions help through directional, selectivity, budget-matched, and held-out perturbation controls. These analyses connect observable spatial instability to testable correction behavior, while explicitly distinguishing diagnostic evidence from a reliable predictor of individual correction gains.

  2. Demonstrating benefits beyond conventional multi-view aggregation. We added controlled comparisons using the same six-view predictions and output budget, including NMS, Soft-NMS, and weighted boxes fusion. Under the matched candidate-threshold setting, FSCI achieves 42.667 AP on the complete COCO val2017 set, compared with 39.042 AP for the strongest evaluated conventional aggregation control. A separately evaluated, approximately latency-matched high-resolution baseline further tests whether the improvements can be explained simply by additional computation or image detail.

  3. Establishing applicability across detector generations. The revised evaluation covers seven detectors on COCO, TinyPerson, and VisDrone2019, including DEIMv2-S and RT-DETRv3-R50. FSCI consistently improves overall AP and small-object AP across all 21 detector–dataset combinations without updating model weights or selecting dataset-specific intervention parameters. Scale-specific component analyses additionally clarify complementarity and remaining cross-scale trade-offs rather than claiming unconditional preservation of large-object performance.

  4. Making the practical scope and cost explicit. We added per-detector latency, throughput, and memory measurements, together with component-level operating points. The full method incurs approximately 5.9–6.2× inference latency in our implementation, while the two-view correction component offers a lower-cost alternative. We also clarified the relationship to training-free test-time adaptation, including ViTPrompt and FACTOR, and positioned our contribution around geometric localization correction, ownership-based candidate handling, and controlled mechanistic validation.

Overall, the revised manuscript presents both a practical training-free intervention and an experimentally examined account of its spatial correction behavior. We thank you for your consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.