acceptodds
Under review as a conference paper at ICLR 2027

GateSlice: Scale-Selective Sliced Fusion for Training-Free Small-Object Detection with Frozen Detectors

Abstract

A fundamental limitation of frozen visual predictors is that potentially useful evidence may become inaccessible when the test input is mapped to the model's fixed operating resolution. For object detection, globally resizing a high-resolution image severely compresses small objects, whereas locally enlarging image regions restores their visual evidence but induces a different prediction distribution because of incomplete context, object truncation, and cross-view confidence shifts. This gives rise to a general inference-time question: how can a frozen model exploit complementary transformed views without allowing less reliable views to destabilize its original predictions? We address this problem through scale-conditioned selective fusion and propose GateSlice, a dual-view, fine-tuning-free framework for small-object detection. GateSlice is based on the observation that view reliability is scale dependent. The global view preserves complete scene context and provides stable predictions, particularly for medium and large objects, while tiled views improve the effective resolution of small objects but are less reliable outside the small-object regime. Accordingly, GateSlice does not treat the two views as equally weighted ensemble members. Instead, it assigns them asymmetric roles: the global branch serves as the prediction anchor, whereas the tiled branch acts as a scale-restricted source of complementary evidence. After tiled detections are mapped to the original image coordinates, an area gate selectively routes only small-scale candidates into fusion. A detector-specific fixed score factor aligns cross-view confidence levels, and class-aware non-maximum suppression resolves duplicate predictions. Together, these designs constrain the influence of the transformed view while preserving useful local evidence. Experiments on COCO, TinyPerson, and VisDrone2019 show that GateSlice consistently improves small-object detection across five frozen detectors, including the transformer-based RT-DETR and Deformable DETR, the one-stage YOLO11n and FCOS, and the two-stage Faster R-CNN. These improvements are obtained without introducing learnable parameters or updating model weights. Multifactor ablations and instance-level transition analyses further show that the gains predominantly arise from recovering small objects missed under global resizing, while scale gating and confidence calibration restrict the interference of tiled predictions with the global results. The consistent gains across distinct detection paradigms indicate that GateSlice is not tied to a particular architecture. Instead, inference-time enhancement of frozen detectors can be formulated as reliability-aware evidence selection across complementary views rather than unconstrained prediction aggregation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Small Object Detection, Training-Free Inference, Sliced Inference, Cross-View Fusion, Scale-Aware Gating

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

The paper introduces a training-free approach aimed at improving small-object detection in high-resolution images. The proposed method can be viewed as an inference-time strategy that complements the conventional full-image prediction. In addition to processing the entire image, the method divides the image into overlapping tiles, with each tile producing additional detections. The resulting predictions are then filtered so that only small objects with an area below 1024 pixels are retained. Their confidence scores are further rescaled using a hyperparameter, after which non-maximum suppression (NMS) is applied to remove duplicate detections. The authors evaluate the method across five different object detectors and three datasets, reporting consistent improvements in detection performance.

Critical strengths
  1. The proposed method follows a training-free paradigm and can be easily integrated with existing object detectors to improve small-object detection performance.
  2. The paper provides empirical validation across five object detectors and three datasets, demonstrating the general applicability of the approach.
  3. The authors conduct ablation studies to assess the contribution of the different components of the proposed method.
Critical weaknesses
  1. Lack of novelty compared to prior works mentioned below since most of the papers idea have been already covered such as global and local processing, overlapping image slices, aggregation of detections, and NMS.

Fatih Cagatay Akyon, Sinan Onur Altinuc, and Alptekin Temizel. “Slicing Aided Hyper Inference and Fine-Tuning for Small Object Detection.” 2022 IEEE International Conference on Image Processing (ICIP), pp. 966–970, 2022. DOI: 10.1109/ICIP46576.2022.9897990.

Xin Yang, Yong Song, Ya Zhou, Yizhao Liao, Jinqi Yang, Jinxiang Huang, Yiqian Huang, and Yashuo Bai. “An Efficient Detection Framework for Aerial Imagery Based on Uniform Slicing Window.” Remote Sensing, 15(17), 4122, 2023. DOI: 10.3390/rs15174122.

  1. The choice of critical hyperparameters is not well motivated such as confidence scalar, windows size and size of object (what has been considered small in COCO dataset may not necessarily justify we need to only keep detections with such size and it may not be relevant for other datasets)

  2. While the approach is training-free, the Reproducibility section is missing.

  3. Section 4.4 is extended yet lacks clarity and coherency.

  4. Figure 5 is provided as extra evidence to compare global and local view. However, the cited appendix shows that the maps calculate only feature value independent of object class or positive/negative contribution.

Questions for authors
  1. What is the specific contribution beyond existing sliced inference and scale-selective fusion? SAHI already supports inference with pretrained detectors without fine-tuning, while Yang et al. (2023) combines global/local detection with scale filtering. Please identify which aspect of GateSlice is not covered by these approaches and explain why it constitutes a substantive methodological or empirical advance.

SAHI: Fatih Cagatay Akyon, Sinan Onur Altinuc, and Alptekin Temizel. “Slicing Aided Hyper Inference and Fine-Tuning for Small Object Detection.” 2022 IEEE International Conference on Image Processing (ICIP), pp. 966–970, 2022. DOI: 10.1109/ICIP46576.2022.9897990.

  1. Does GateSlice outperform these closely related methods under controlled conditions?

  2. How were the detector-specific confidence multipliers selected, and do they transfer without retuning?

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

GateSlice is a training-free inference framework for improving small-object detection with frozen detectors. It applies the same frozen detector to a full-image view and overlapping tiled views, maps tiled predictions back to the original image coordinates, filters them using a fixed area threshold, and adjusts their confidence scores using a detector-specific fixed factor. The retained tiled predictions are then fused with the full-image predictions through class-aware NMS and final Top-K selection. The method is evaluated on COCO val2017, TinyPerson, and VisDrone2019 using five COCO-pretrained base detectors.

Critical strengths
  1. Consistent improvements across detectors.
    GateSlice improves all five evaluated frozen detectors, covering one-stage, two-stage, and Transformer-based architectures.
  2. Focused ablations and instance-level analysis.
    The ablations isolate the effects of area filtering, score scaling, and NMS, while the instance-level analysis quantifies both rescued and lost detections. Together, they show how constrained fusion recovers small objects while limiting interference from tiled predictions.
Critical weaknesses
  1. Limited differentiation from prior work.
    SAHI already supports detector-agnostic sliced inference without additional fine-tuning, including overlapping crops, optional full-image inference, coordinate restoration, and NMS-based fusion. SNIP filters detections by scale-specific validity ranges during inference before Soft-NMS fusion, although its full method also involves specialized training. GateSlice's main distinctions are its asymmetric branch design, fixed area gate, and detector-specific score scaling. Although Table 4 supports the utility of these components within the proposed pipeline, their contribution relative to existing sliced and scale-selective inference methods remains insufficiently established.
  2. Insufficient comparative and computational evaluation.
    Tables 1–3 compare GateSlice with standard full-image inference. Table 4 includes a useful SAHI-like symmetric NMS control, but its correspondence to SAHI is not explained, and the comparison covers only Deformable-DETR-R50 on COCO. A documented comparison with SAHI using full-image inference under controlled settings is needed to assess the benefit over existing sliced inference. The default configuration processes one global view and four tiled views, yet neither end-to-end latency nor peak memory is reported, leaving the accuracy–cost trade-off unclear. The selection procedure for the detector-specific score factors is also unspecified. Moreover, the marginal score distributions in Fig. 4(b) motivate investigating cross-view score differences but do not establish that the proposed scaling improves confidence calibration.
  3. How does GateSlice differ substantively from SAHI and SNIP-style scale-selective inference? Is the “Symmetric NMS fusion” in Table 4 intended as a SAHI baseline? Please clarify this correspondence and provide comparisons with SAHI and a frozen, inference-only scale-filtering baseline across detectors and datasets under matched settings.
  4. What are the computational costs and parameter-selection protocol of GateSlice? Please report end-to-end latency and peak memory relative to full-image and sliced-inference baselines, and explain how the detector-specific score factors were selected and whether they were fixed before cross-domain evaluation.
Questions for authors

The authors need to respond to weaknesses in their rebuttal.

Reviewer_zr3wOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GateSlice, a method for improving small object detection without retraining. It uses the same frozen detector to process both the full image and overlapping tiles, with tiling increasing the effective size of objects in the network input. The tiled predictions are then mapped back to the original image coordinates. An area threshold retains small object candidates, and a fixed factor adjusts their confidence scores before they are fused with full-image predictions using NMS. The authors evaluate the method across five general-purpose detectors on COCO, TinyPerson, and VisDrone, reporting improvements in small object detection accuracy and recall.

Critical strengths

The paper is reasonably well written, and the proposed method improves small object detection accuracy across the general-purpose detectors evaluated in the paper.

Critical weaknesses
  1. The differences from existing sliced inference methods are not sufficiently clear. Sliced training and inference have long been used for small object detection. SAHI [1] already supports sliced inference without fine-tuning, as well as the fusion of full-image and sliced predictions. GateSlice’s additional design mainly consists of filtering sliced detections by bounding-box area, adjusting their confidence scores using a fixed factor, and applying NMS. These rules are relatively straightforward, and the paper does not sufficiently explain what substantive technical advances they offer over existing methods.

  2. Comparisons with directly related methods and strong small object detection baselines are missing. Tables 1 to 3 mainly compare five general-purpose detectors with their GateSlice-enhanced versions. These results demonstrate effectiveness across different detectors but are insufficient to establish an advantage over existing methods. In particular, SAHI [1], which closely matches the setting of this paper, is cited but not directly compared. Other representative methods, including QueryDet [2], CEASC [3], ESOD [4], and YOLC [5], are also absent from the experiments.

  3. Inference speed evaluation is missing. The default configuration processes one full image and four overlapping tiles, followed by prediction fusion. On COCO, the paper reports improvements of approximately 0.4 to 1.2 points in overall AP and 1.1 to 2.7 points in small object AP, but provides no measurements of end-to-end latency, FPS, or computational overhead relative to the original detectors.

  4. The practical motivation and applicability of the training-free design are insufficiently clarified. The paper emphasizes that deployed detectors require neither modification nor retraining. However, it does not adequately explain which applications cannot accommodate one-time fine-tuning but can tolerate additional inference computation for every input. When annotations and training resources are available, a modest amount of fine-tuning is also a reasonable option. The authors should therefore clarify the practical constraints that require keeping the detector frozen and explain the deployment benefits of their approach relative to other alternatives.

References

[1] Fatih Cagatay Akyon, Sinan Onur Altinuc, and Alptekin Temizel. Slicing Aided Hyper Inference and Fine-tuning for Small Object Detection. ICIP, 2022.

[2] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. QueryDet: Cascaded Sparse Query for Accelerating High-Resolution Small Object Detection. CVPR, 2022.

[3] Bowei Du, Yecheng Huang, Jiaxin Chen, and Di Huang. Adaptive Sparse Convolutional Networks with Global Context Enhancement for Faster Object Detection on Drone Images. CVPR, 2023.

[4] Kai Liu et al. ESOD: Efficient Small Object Detection on High-Resolution Images. IEEE Transactions on Image Processing, 34: 183–195, 2025.

[5] Chenguang Liu et al. YOLC: You Only Look Clusters for Tiny Object Detection in Aerial Images. IEEE Transactions on Intelligent Transportation Systems, 25(10): 13863–13875, 2024.

Questions for authors

refer to Critical Weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.