acceptodds
Under review as a conference paper at ICLR 2027

QAD-D-FINE: QUERY-ADAPTIVE SAMPLING AND REDUNDANCY SUPPRESSION FOR DENSE SMALL OBJECT DETECTION

Abstract

Dense small object detection is challenging because weak fine-grained cues and crowded layouts jointly cause missed objects and redundant responses. We propose QAD-D-FINE, a query-adaptive, redundancy-suppression extension of D-FINE for dense small-object scenes. QAD-D-FINE contains two complementary components. Query-Adaptive Sampling (QAS) augments the decoder with a high-resolution P2 feature and continuously adjusts deformable sampling offsets according to the incoming query scale and confidence, allowing confident small-object queries to aggregate more local evidence while preserving context for uncertain queries. Matching-based Redundant Query Suppression (MRS) uses the standard one-to-one matching result as an anchor, assigns unmatched queries to same-class ground-truth instances with geometry-normalized ownership, and jointly suppresses redundant classification and localization responses during training. On VisDrone, QAD-D-FINE reaches 33.1 AP and 23.6 APsmall, improving D-FINE by 1.7 and 3.3 points, respectively, while retaining 166.7 FPS with 10.1M parameters at 640 × 640 input. Experiments on SODA-D and UAVDT further demonstrate consistent gains in dense small-object scenarios.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Dense Small Object Detection, Query-Adaptive Sampling, Redundant Query Suppression, Deformable Attention

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_Ki48OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces QAD-D-FINE to improve the detection of dense, small-scale objects. It employs two key mechanisms: Query-Adaptive Sampling (QAS), which incorporates P2 features and dynamically scales deformable sampling offsets based on query-specific scales and confidence levels and Matching-based Redundant Query Suppression (MRS), a loss function applied during training to penalize redundant responses. QAD-D-FINE achieves consistent performance gains on the VisDrone, SODA-D, and UAVDT datasets with an increase of only about 0.1M parameters.

Critical strengths

(1) MRS suppresses redundancy at zero inference cost. It operates only during the training phase on unmatched queries, leaves the core one-to-one matching mechanism intact, and does not enter the inference graph.

(2) Randomly permuting the \gamma factors for each query within the same image preserves the marginal distribution of \gamma, directly demonstrating that the gains stem from the "correspondence between the scaling factor and the query state," rather than from shift normalization or changes to the global sampling radius.

Critical weaknesses

(1) Adding a P2 high-resolution branch to a detector to handle small objects has become a standard practice for improving YOLO/DETR-based models since 2020; as early as 2024, DQ-DETR [1] incorporated five-scale feature maps into a DETR-style architecture for the detection of tiny objects. QAS primarily enables the sampling range to vary according to the query state. However, BoxR [2] has proposed query conditioning and scaling the sampling range in proportion to the box scale. Therefore, the contribution of QAS is minimal and incremental.

(2) Using the Hungarian matching results as an anchor, the MRS method of applying "soft negative sample suppression" to unmatched queries of the same type is essentially equivalent to assigning soft labels or designating "ignore regions" (similar to the approaches in FreeAnchor [3] or ATSS [4]) for negative samples that lie close to positive samples.

(3) The core motivation behind MRS is that unmatched queries (following one-to-one matching) tend to cluster around already-matched targets, leading to redundant responses. It is precisely the issue addressed by DDQ [5]. Furthermore, "one-to-many" auxiliary assignment approaches (such as Group DETR [6], H-DETR [7], etc.) are conceptually related to the "matching relationship reuse" mechanism in MRS. However, these works are entirely absent from the "Related Work" section, and no comparative experiments involving them have been conducted.

(4) The motivation behind introducing P2 in QAS is to mitigate the loss of detail in small objects caused by downsampling. However, the standard practice for addressing this same issue is simply to increase the input resolution. While the paper compares QAS@640 with D-FINE@640, it fails to provide performance curves for D-FINE at higher resolutions, such as 768 or 896. Table 2 indicates that QAS comes at the cost of an 18% increase in latency (from 5.1 ms to 6.0 ms). This is on the same order of magnitude as the cost incurred by moderately increasing the input resolution.

(5) The score, defined as \text{IoU} + \exp(-q), raises several questions: Why are the distance term and the IoU term added rather than multiplied? Why is this mechanism applied only to ground-truth (GT) objects that are small (and why is there no need to suppress redundant queries for large objects)? Furthermore, why is it applied exclusively to the final decoder layer, whereas D-FINE’s auxiliary losses are distributed across all layers? None of these design choices have been subjected to an ablation study.

[1] DQ-DETR: DETR with Dynamic Query for Tiny Object Detection

[2] BoxeR: Box-Attention for 2D and 3D Transformers

[3] FreeAnchor: Learning to Match Anchors for Visual Object Detection

[4] Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection

[5] Dense Distinct Query for End-to-End Object Detection

[6] Group DETR: Fast DETR Training with Group-Wise One-to-Many Assignment

[7] DETRs with Hybrid Matching

Questions for authors

[1] What are the primary innovations of this paper? The authors should elaborate on them in detail and provide sufficient experimental evidence to support their claims.

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

This paper proposes QAD-D-FINE, an extension of D-FINE for dense small-object detection. Query-Adaptive Sampling (QAS) incorporates the high-resolution P2 feature and adjusts deformable sampling offsets according to query scale and confidence. Matching-based Redundant Query Suppression (MRS) uses Hungarian matching results to identify unmatched queries and suppress redundant classification and localization responses during training. Experiments on VisDrone, SODA-D, and UAVDT, supported by ablation studies, demonstrate consistent improvements over D-FINE, including +1.7 AP and +3.3 APs on VisDrone. However, the paper does not sufficiently differentiate its methods from closely related works or provide empirical evidence supporting the claimed sampling behavior of QAS.

Critical strengths
  1. QAD-D-FINE achieves improvements over D-FINE across VisDrone, SODA-D, and UAVDT, demonstrating the effectiveness of the proposed approach across multiple dense small-object detection benchmarks.

  2. The paper provides ablation studies examining the individual contributions of QAS and MRS, as well as several design choices within each module. These experiments help establish the contribution of the proposed components to the overall performance improvements.

Critical weaknesses
  1. Insufficient differentiation from closely related work. QAS shares the idea of adjusting sampling based on query scale with AdaMixer [1], while MRS addresses redundant queries, a problem already explored by DDQ [2] through distinct query selection. However, the paper does not sufficiently clarify the methodological advantages of QAS and MRS over these existing approaches. Moreover, Dome-DETR [3], which also builds on D-FINE for dense small-object detection, is absent from the experimental comparisons. Without a comparison under consistent evaluation settings, it is difficult to assess the relative advantages and significance of the proposed contributions.

  2. Lack of evidence supporting the claimed sampling behavior of QAS. The original D-FINE already scales deformable sampling offsets according to the reference box width and height, whereas QAS replaces this rule with feature map size normalization and a scale- and confidence-dependent factor. Consequently, a value of γ\gamma below 1 does not necessarily imply more localized sampling than the original D-FINE, particularly for small objects. Although the Q0–Q5 ablations demonstrate performance improvements, they do not directly verify the claimed sampling behavior. An analysis of the actual sampling distances across object scales and feature levels would help establish whether QAS operates as described.

Reference

[1] Gao, Ziteng, et al. "Adamixer: A fast-converging query-based object detector." 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022.

[2] Zhang, Shilong, et al. "Dense distinct query for end-to-end object detection." 2023 IEEE/CVF conference on computer vision and pattern recognition (CVPR). IEEE, 2023.

[3] Hu, Zhangchi, et al. "Dome-DETR: DETR with density-oriented feature-query manipulation for efficient tiny object detection." Proceedings of the 33rd ACM International Conference on Multimedia. 2025.

Questions for authors
  1. Clarify the methodological differences between QAS and AdaMixer [1], which also adapts sampling locations and feature scale to each query. Do the same for MRS and DDQ [2], which also addresses redundant queries in one-to-one end-to-end detection. Support the claimed advantages empirically, for example by replacing QAS with an AdaMixer-style scale-adaptive sampling rule and MRS with DDQ-style distinct query selection within the same D-FINE-S framework. Provide a direct comparison with Dome-DETR [3] under identical input resolution, data split, and evaluation protocol

  2. Since the original D-FINE already scales sampling offsets according to the reference box size, it is unclear whether QAS actually produces more localized sampling for small objects. An analysis of the actual sampling distances across object scales and feature levels for Q0, Q1, and Q4 would help clarify this distinction and better support the claimed sampling behavior.

  3. Provide evidence linking the actual sampling behavior of QAS to the improvements in APs and ARs. If QAS produces broader sampling than the original D-FINE for small objects, explain how this behavior leads to the reported performance gains and reconcile it with the claimed mechanism.

Reviewer_2D3iOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problems of fixed decoder sampling ranges and insufficient suppression of unmatched redundant queries in dense small-object detection, and proposes the QAD-D-FINE framework. QAS introduces P2 high-resolution features into the decoder and continuously scales deformable sampling offsets according to the query scale and confidence entering the current layer; MRS uses Hungarian matching during training to suppress unmatched queries that have high class responses and are geometrically redundant. The authors conduct main experiments, module ablations, efficiency tests, multi-seed experiments, and hyperparameter sensitivity analyses on VisDrone, SODA-D, and UAVDT. On VisDrone, the method improves over D-FINE by 1.7 AP, 3.3 APs, and 7.7 ARs, with 10.1M parameters, 28.0 GFLOPs, and 166.7 FPS on an RTX 4090.

Critical strengths

The proposed method introduces P2 high-resolution sampling with only a small increment in parameters and FLOPs, and continuously couples the sampling range with query scale and confidence. It does not alter the Hungarian primary matching or the inference graph, thereby suppressing redundant responses with zero inference overhead. At a relatively small computational cost, the method improves APs and ARs on VisDrone by 3.3 and 7.7, respectively, and reduces the duplicate prediction rate; the gains are mainly concentrated on small-object recall rather than being uniform across all scales.

Critical weaknesses
  1. Small objects are a pain point in visual detection tasks, especially for DETR-series methods, and queries are the main factor affecting small-object performance. In discussing this issue, the authors did not include the latest methods. Backbone is a relatively critical component here; since different methods use different backbones, discussing them together directly makes it difficult to directly address the authors’ main innovation.

  2. Small-object detection is fundamentally a problem of information deficiency and feature separability. This paper only alleviates two superficial symptoms—sampling and redundancy—rather than fundamentally improving the separability of small objects.

Questions for authors
  1. Small-object detection in real deployment often involves illumination changes, rain, fog, and motion blur. This paper is evaluated only on the standard VisDrone, SODA-D, and UAVDT benchmarks, without degradation-augmentation experiments such as cloud/fog simulation, downsampling, or motion blur. Can the model still maintain its gains when image quality degrades or targets are severely occluded?

  2. QAS uses predicted box area and maximum class probability to determine the sampling range, but this mapping lacks a formal basis in terms of effective receptive field, sampling density, localization uncertainty, or signal-to-noise ratio. Low confidence may arise from occlusion, small object size, early training, or annotation noise; is it reasonable to approach fixed sampling in such cases?

  3. Although evaluated on SODA-D and UAVDT, the paper still focuses on aerial and traffic-surveillance scenes, and it does not specify whether the cross-dataset experiments involve separate training per dataset or direct transfer. Can the model generalize to cross-city, cross-altitude, or ground-driving scenarios? Can the authors add direct-transfer or few-shot adaptation experiments and clarify the effects of scale-prior and matching-distribution shifts?

  4. QAS uses the same γ to scale the normalized offsets of P2–P5, but at different resolutions the same γ corresponds to inconsistent physical displacements; P2 has weak low-level semantics and large noise, and changes in sampling positions may also alter the gradient paths of D-FINE distribution refinement and reference-box updates. MRS suppresses unmatched high-response queries, which may affect the stability of the next-round Hungarian matching and the assignment of positive and negative samples. Can the authors analyze the error propagation and sampling-geometry consistency in these intermediate stages?

  5. Ablations show that P2 alone brings clear gains; QAS improves over permuted γ by only about 0.5 AP, and MRS is a training-time regularizer; Full improves over +QAS by only about +0.6 AP. Is the innovation increment of QAS and MRS sufficient, or do the gains mainly come from P2 and training regularization? Can the authors add multi-seed results for +QAS alone, multi-seed cross-dataset results, and more direct causal experiments?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.