acceptodds
Under review as a conference paper at ICLR 2027

Reasoning-DETR: Contextual and Relational Query Learning for Aerial Object Detection

Abstract

Object detection has progressed rapidly from two stage and one stage frameworks to Transformer detectors. Among them, DETR formulates detection as set prediction and allows object queries to exchange contextual information through attention. However, its direct application to aerial imagery remains limited. Tiny objects are easily suppressed by complex backgrounds, while their number and spatial distribution vary considerably across scenes. A fixed query budget therefore cannot accommodate different search demands, and interactions based on feature similarity may introduce unreliable background or neighboring evidence into ambiguous queries. To address these issues, we propose Reasoning-DETR, which couples adaptive object search with query reasoning based on uncertainty. First, a dynamic query allocator guided by scene context combines regional density prediction with Scene Memory to determine the number and spatial distribution of supplementary queries. Second, the query reasoner retrieves category prototypes for ambiguous queries and selectively aggregates evidence from reliable nearby objects. A reliability gate suppresses relational messages that conflict with the current query representation. Both modules are integrated into the original RT-DETR pipeline without requiring additional scene or relation annotations. Experiments on VisDrone and AI-TOD-V2 demonstrate consistent improvements in aerial and tiny-object detection, with improved query coverage and recognition of weak objects at a modest computational overhead.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Aerial object detection, Detection Transformer, Large language models, Relational reasoning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Rci6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Reasoning-DETR, an RT-DETR-based framework for aerial object detection, targeting the challenges of tiny, densely distributed, and visually ambiguous objects. The method contains two main components. First, a Scene-Aware Dynamic Query Allocation module adaptively introduces supplementary queries according to regional density estimates and a prototype-based Scene Memory, aiming to improve query coverage in scenes with varying object densities. Second, an Uncertainty-Aware Semantic and Relational Query Reasoner identifies ambiguous supplementary queries, refines them using category prototypes, and selectively aggregates information from reliable nearby queries through gated relational interaction. Experiments on VisDrone and AI-TOD-V2 show consistent improvements over RT-DETR and several aerial detection baselines, particularly for tiny objects, while introducing a moderate increase in computational complexity. Overall, the paper presents a coherent framework that combines adaptive query allocation with contextual query refinement for aerial detection.

Critical strengths
  • The paper is clearly organized, and the proposed framework is presented in a structured manner.
  • The experimental results are generally solid, showing consistent improvements over the RT-DETR baseline and competitive performance on both VisDrone and AI-TOD-V2.
Critical weaknesses
  • The key motivation behind the proposed query reasoner is insufficiently validated. The paper assumes that ambiguous small object queries are harmed by misleading background or neighboring-query interactions, but provides no direct analysis showing that this failure mode actually occurs or that the proposed module specifically alleviates it. As a result, the motivation remains largely intuitive rather than empirically established.

  • The Scene-Aware Dynamic Query Allocation module has limited novelty and relies on several heuristic, dataset-dependent design choices:

    1. The module is closely related to prior density-guided dynamic-query approaches such as DQ-DETR. Its main additional design is to partition the high-resolution feature map into overlapping regions and perform query budgeting at the regional level. Therefore, the contribution appears to mainly refine the spatial granularity of existing dynamic-query allocation rather than introducing a fundamentally new query generation or allocation mechanism.
    2. The allocation mechanism contains many heuristic design choices and manually specified hyperparameters. The proposed pipeline introduces a considerable number of hyperparameters, including the regional partition parameter MM, overlap ratio ρ\rho, Scene Memory size KsK_s, number of retrieved prototypes LL, number of density levels RR, predefined density-to-budget mappings {Br}\{B_r\}, Scene Memory weighting coefficient λs\lambda_s, and the maximum number of supplementary queries Kmax⁡K_{\max}. In particular, the regional query demand formulation is largely hand-crafted.
    3. Scene Memory may overly rely on dataset-specific contextual biases. Scene Memory is constructed exclusively from training regions containing ground-truth objects, meaning that the resulting prototypes largely encode object-context co-occurrence patterns observed in the training set. Such priors can be effective under an IID evaluation protocol, but their robustness under scene or domain shifts remains unclear.
  • The Uncertainty-Aware Semantic and Relational Reasoning module have several insufficiently justified design choices:

    1. The individual components are largely built upon standard techniques, including entropy-based uncertainty estimation, classifier weights as category prototypes, prototype attention, spatial nearest-neighbor selection, geometric relation bias, and gated feature aggregation. Therefore, the contribution mainly lies in combining these existing mechanisms for aerial query refinement rather than introducing a fundamentally new reasoning formulation.
    2. Several components are heuristic. The method normalizes sigmoid classification scores to construct a categorical distribution for entropy estimation, defines neighbor reliability as ρj=(1−uj)max⁡csj(c)\rho_j=(1-u_j)\max_c s_j(c), filters queries using a manually specified threshold τr\tau_r, and further selects a fixed number of eight spatial neighbors. The rationale and robustness of these choices are not sufficiently justified, and their sensitivity should be more thoroughly evaluated.
    3. The term ``reasoning'' appears somewhat overstated for the proposed mechanism. In practice, the module mainly performs uncertainty-based query selection, prototype retrieval, local neighbor aggregation, and gated feature fusion, all of which are relatively standard representation refinement operations.
  • The experimental evaluation has several limitations:

    1. The actual inference efficiency is not sufficiently evaluated. Although the paper reports Params and GFLOPs, it does not provide FPS or inference latency. This is particularly important because the proposed method introduces dynamic supplementary queries, prototype retrieval, neighbor selection, and additional attention operations, whose practical runtime overhead may not be fully reflected by GFLOPs. Reporting FPS under a clearly specified hardware would make the efficiency claim more convincing.
    2. The comparison with recent detectors is not sufficiently comprehensive. Several baselines in the main table are relatively old, while stronger and more recent real-time detectors are missing. For example, the paper does not include newer YOLO variants such as YOLO11, DEIM, RT-DETRv4, nor other recent competitive detectors under comparable training and input-resolution settings.
  • The period at line 394 should be carefully revised.

Questions for authors

See Weakness.

6 comments

Thank you for your constructive feedback and for recognizing our work. We considered the concerns and conducted additional analyses to validate our method. We hope our responses clarify its contributions and improvements, and appreciate your consideration of a higher score.

Weakness 1: The motivation behind the proposed query reasoner is insufficiently validated.

A: To directly investigate the influence of unreliable neighboring information, we conducted additional controlled experiments.

(1) We compared semantic refinement, unfiltered nearest-neighbor aggregation, reliability filtering, and reliability filtering with the proposed gate.

(2) As shown in Table 1, unfiltered aggregation introduces 38.5% background neighbors, resulting in lower query accuracy (44.9%) and higher NLL (1.33).

(3) Reliability filtering and gating improve query accuracy to 50.2%, reduce NLL to 1.14, and achieve 33.5 AP, demonstrating their effectiveness in suppressing misleading neighboring information.

Table 1. Analysis of unreliable neighboring information on VisDrone.

Neighbor StrategyFG Neighbors (%)BG Neighbors (%)APAccu.NLL
Semantic Prototypes Only––32.747.01.25
Unfiltered Nearest-861.538.532.344.91.33
Reliability Filter, w/o Gate82.018.033.148.61.20
Reliability Filter + Gate82.018.033.550.21.14

Weakness 2: The Scene-Aware Dynamic Query Allocation module has limited novelty and relies on heuristic, dataset-dependent designs.

A: We further clarify the distinction from DQ-DETR and Dome-DETR. Unlike methods primarily relying on image-dependent density information, our method introduces Scene Memory as an additional contextual prior.

(1) As shown in the original ablation, increasing Top-K queries from 300 to 500 improves AP from 29.7 to 30.7; adding Scene Memory raises it to 31.2, and incorporating density prediction further improves it to 32.0.

(2) GT coverage also improves from 90.8% to 91.7%, with small-object coverage increasing from 84.3% to 86.1% (Table 7).

To examine the heuristic designs, we conducted sensitivity analyses on representative parameters, as shown in Tables 2–4.

Table 2. Sensitivity analysis of reliability threshold and scene-memory weight on UAVDT.

τ \ α0.250.500.751.00
0.3024.725.024.924.5
0.4025.125.525.425.0
0.4525.325.825.625.2
0.5025.025.525.324.9
0.6024.625.024.824.4

Table 3. Sensitivity analysis of supplementary query budget on VisDrone.

KmaxMaximum Total QueriesAPAP50GFLOPs
030029.750.6108.0
10040030.451.2110.2
20050031.751.6112.3
30060032.652.9114.4
50080033.554.2118.0
700100033.654.3122.1

Table 4. Sensitivity analysis of representative hyperparameters on VisDrone.

HyperparameterTested ValuesCorresponding AP
Region Grid3×3, 4×4, 5×5, 6×633.0, 33.3, 33.5, 33.2
Overlap Ratio0, 0.15, 0.25, 0.35, 0.5033.1, 33.3, 33.5, 33.4, 33.0
Memory Size16, 32, 64, 12833.2, 33.4, 33.5, 33.3
Retrieved Prototypes1, 2, 4, 833.2, 33.4, 33.5, 33.2
Density Levels3, 4, 5, 633.1, 33.4, 33.5, 33.2
Neighbor Number4, 6, 8, 10, 1233.2, 33.4, 33.5, 33.3, 33.0

These results show that the selected configurations perform consistently well. Other settings, such as density-to-budget mappings, are coupled with the overall allocation policy and are therefore evaluated through the corresponding budget analysis. We will provide complete default configurations for reproducibility.

We also acknowledge that Scene Memory may reflect training-distribution characteristics and will clarify its potential limitations under domain shifts.

Weakness 3: The Uncertainty-Aware Reasoning module relies on standard techniques and insufficiently justified heuristic choices.

A: Our contribution lies in integrating semantic evidence and reliable local relations into a structured query refinement process, rather than introducing each underlying operation independently.

(1) Category prototypes provide semantic guidance, while reliable neighbors supply complementary local information through gated aggregation.

(2) As shown in Table 8, progressive refinement improves AP from 32.0 to 32.7, 33.1, and 33.5, while uncertain-query accuracy increases from 42.1% to 50.2% and NLL decreases from 1.39 to 1.14.

Regarding the term Reasoning, we use it to describe structured semantic and relational evidence refinement at the query level, rather than multi-step symbolic reasoning. We will clarify this definition to avoid overstating the method's scope.

Additional Responses to the Reviewer

Weakness 4: The actual inference efficiency is insufficiently evaluated.

A: We supplemented the evaluation with inference latency and FPS measurements on a single NVIDIA RTX 3090 GPU, using 640 × 640 inputs, FP16 precision, and batch size 1.

Table 5. Inference efficiency comparison on RTX 3090.

ModelParams (M)GFLOPsLatency (ms)FPS
RT-DETR-L32.9108.016.062.5
+ Adaptive Query Allocation33.5116.517.557.1
Reasoning-DETR34.0118.018.254.9

Although supplementary-query decoding and refinement introduce additional overhead, Reasoning-DETR maintains real-time inference speed while improving AP from 29.7 to 33.5. We will include these efficiency measurements and the hardware configuration in the revised manuscript.

Weakness 5: The comparison with recent object detectors is not sufficiently comprehensive.

A: We expanded the comparisons to include YOLO11, DEIM, and RT-DETRv4, using the same VisDrone split and 640 × 640 input resolution.

Table 6. Comparison with recent object detectors on VisDrone.

MethodInput SizeAPAP50
YOLO11-L640 × 64030.851.8
DEIM-D-FINE-L640 × 64031.752.9
RT-DETRv4-L640 × 64031.452.4
RT-DETR-L640 × 64029.750.6
Reasoning-DETR (Ours)640 × 64033.554.2

The results demonstrate the competitiveness of our method against recent real-time detectors.

Weakness 6: The period at line 394 should be carefully revised.

A: Thank you for identifying this issue. We have corrected the punctuation and further checked the manuscript for similar typographical and formatting errors.

Dear Reviewer Rci6,

Thank you again for your constructive suggestions. Following our previous responses in the rebuttal, we have further revised the manuscript by incorporating additional experimental analyses, including inference efficiency evaluations and hyperparameter sensitivity studies. We have also carefully checked the manuscript for typographical and formatting issues. We sincerely hope these revisions address your concerns and further strengthen our work. The corresponding changes are highlighted in purple.

Reviewer_3fVzOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Reasoning-DETR for aerial object detection. It combines regional density estimates with Scene Memory to allocate supplementary queries, then refines uncertain queries using category prototypes and reliable neighboring queries. Built on RT-DETR, the method requires no additional scene or relation annotations. Experiments on VisDrone and AI-TOD-V2, together with component ablations, evaluate its detection performance and the contributions of the proposed modules.

Critical strengths

The method is well motivated by practical challenges in aerial object detection. Its design is reasonable, and it achieves clear performance improvements on established aerial detection benchmarks.

Critical weaknesses

See Questions.

Questions for authors
  1. DQ-DETR [1] and Dome-DETR [2] have already explored density-guided dynamic query allocation. Could the authors clarify the contribution beyond these methods and explain why Scene Memory provides additional benefits over regional density prediction alone? Could the authors provide visualizations to more clearly illustrate the role of Scene Memory?

  2. The method retains 300 base queries and dynamically adjusts only the number of supplementary queries. Since aerial imagery also includes sparse scenes, could the authors explain why the number of base queries is fixed and whether they have explored adapting it to scene density?

  3. Background queries with low but similar class scores may also have high entropy. Could the authors clarify how uncertain queries are selected, whether foreground confidence is considered, and how such background queries affect the reasoning module?

[1] Huang Y X, Liu H I, Shuai H H, et al. Dq-detr: Detr with dynamic query for tiny object detection[C]//European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024: 290-305.

[2] Hu Z, Wu P, Chen J, et al. Dome-DETR: DETR with density-oriented feature-query manipulation for efficient tiny object detection[C]//Proceedings of the 33rd ACM International Conference on Multimedia. 2025: 101-110.

4 comments

We appreciate the reviewer's thoughtful assessment and encouraging comments on the motivation, design, and experimental performance of our work. Your questions provide an opportunity to further explain several important design choices. We address each point below and hope these clarifications help strengthen your assessment of our contribution.

Question 1: Contribution and visualization of Scene Memory

A: Thank you for this insightful question. Our contribution differs from DQ-DETR and Dome-DETR by incorporating learned scene context into density-guided query allocation, rather than relying solely on regional density estimates.

(1) Specifically, Scene Memory maintains 64 scene-context vectors, updated through momentum aggregation and retrieved to provide complementary contextual guidance for query allocation.

(2) To further verify its contribution, we conducted controlled comparisons under identical decoder and reasoner settings. Density-only achieves 33.0 AP with 672 queries on average, while adding Scene Memory improves performance to 33.5 AP with 684 queries. Table 7 also shows improvements in overall GT coverage (90.8% to 91.7%) and small-object coverage (84.3% to 86.1%).

(3) Fig. 3 visualizes the overall effects of dynamic query allocation. To more clearly illustrate the specific role of Scene Memory, we will additionally provide side-by-side visualizations comparing density-only allocation with Scene Memory guidance.

Question 2: Motivation for fixed base queries

A: Thank you for raising this question. Our design maintains 300 base queries to preserve the global detection coverage of RT-DETR, while dynamically adjusting supplementary queries according to regional object density.

(1) Base queries provide a consistent global detection foundation, whereas supplementary queries enhance coverage in crowded or potentially underrepresented regions. In sparse scenes, fewer supplementary queries are allocated, reducing unnecessary computation.

(2) We have not explicitly investigated dynamically adjusting the number of base queries. Our current design prioritizes stable global coverage while adapting the additional query budget. We will clarify this motivation and discuss joint adaptation as a possible extension.

Question 3: Uncertain-query selection and background interference

A: Thank you for highlighting this important concern. We clarify the distinction between uncertain queries selected for refinement and reliable queries used as supporting information.

(1) We select the 25% of supplementary queries with the highest normalized prediction entropy, as defined in Eq. (8). Foreground confidence is not used as an additional selection criterion at this stage.

(2) For neighboring base queries, reliability is evaluated by combining prediction certainty and maximum class confidence: ρj=(1−uj)max⁡csj(c)\rho_j = (1-u_j)\max_c s_j(c).

Only neighbors satisfying the reliability criterion are considered for relational refinement, while the gating mechanism further regulates their contributions.

(3) We acknowledge that background queries with uniformly low class scores may also exhibit high entropy and thus enter the refinement set. However, such queries are not automatically treated as reliable supporting evidence. The reliability filtering and gating mechanisms are designed to limit the propagation of misleading information. We will make this distinction clearer in the revised manuscript.

Dear Reviewer 3fVz,

Thank you again for your insightful comments. Following our previous responses in the rebuttal, we have further revised the manuscript to clarify the motivation and contribution of Scene Memory, supported by additional experimental analyses. We sincerely hope these revisions address your concerns and further strengthen our work. The corresponding changes are highlighted in green.

Thanks for your detailed responses, which partially solve my concern. I will maintain my score.

Dear Reviewer 3fVz,

We sincerely thank you for your careful consideration and constructive feedback. Your valuable comments have helped us further reflect on our work and improve the quality of our manuscript. We greatly appreciate your time and thoughtful evaluation.

Best regards, Submission 383 Authors

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

This paper extends RT-DETR for aerial detection with two components. A regional density predictor and a scene-prototype memory allocate supplementary queries, while an uncertainty-selected subset retrieves classifier prototypes and gated information from reliable nearby base queries before the last decoder layer. The system uses detection annotations without additional scene or relation labels. It reports 33.5 AP on VisDrone versus 29.7 for RT-DETR, and 36.3 AP on AI-TOD-V2 versus 34.6 for Dome-DETR. The appendix includes five-seed results, query-coverage measurements, component ablations, and a supplementary-budget sweep.

Critical strengths
  1. Useful gains on two relevant benchmarks. The VisDrone improvement is 3.8 AP at 640-pixel input, with reported parameters increasing from 32.9M to 34.0M and computation from 108 to 118 GFLOPs (Table 1, p. 7). The AI-TOD-V2 result improves on the listed Dome-DETR result by 1.7 AP, including 1.4 AP for very tiny objects (Table 2, p. 8). This is meaningful evidence of practical value for small-object detection, although the cross-paper comparisons are not fully controlled.
  2. The appendix provides evidence beyond aggregate AP. Five runs yield 33.50 ± 0.10 AP (Table 6, p. 17). Table 7 reports better pre-decoder small-object coverage, and Table 8 separates semantic refinement, local relations and the gate: 32.0 → 32.7 → 33.1 → 33.5 AP. These analyses support a real contribution from query refinement and make the proposed integration more credible than a single headline comparison.
Critical weaknesses
  1. The allocation comparison does not fully isolate the proposed mechanism. Table 3 compares fixed Top-K totals of 300 and 500 with a method permitting 300 base plus 500 supplementary queries, i.e. up to 800 total (Table 9, p. 18). Table 9 does provide a useful partial control: at a maximum total of 500, the complete method reaches 31.7 AP versus 30.7 for fixed Top-K. However, that comparison also includes the reasoner; Table 7 reports coverage rather than final AP and does not specify matched realized query counts. Final detection AP for global Top-K and density-only allocation at the same query/compute budget would establish the incremental benefit of Scene Memory. This matters because adaptive density-guided query selection is already established by the cited DQ-DETR and Dome-DETR.
  2. The specific value of uncertainty-guided reasoning remains insufficiently separated from extra attention capacity. Table 8 usefully isolates the gate, but does not compare entropy-selected queries with a size-matched random/confidence-selected subset, or the prototype/reliable-neighbor block with an equally sized ordinary attention block. Lower NLL alone also does not establish improved calibration, as claimed in Appendix C. Moreover, Section 3.3 does not state how many queries are selected; the reported VisDrone configuration does not clearly give its reliability threshold or density-loss weight. These omissions affect both reproduction and interpretation of the paper’s main reasoning claim.
Questions for authors
  1. Controlled comparison: At the adopted budget, what are the realized mean/upper-tail query counts and final AP for global Top-K, density-only allocation, and density plus scene memory, keeping the decoder and reasoner identical? If available, include the fixed-800-query baseline or a computation-matched alternative. This would resolve the allocation concern directly.
  2. Mechanism clarification: What exact rule and count/fraction define the uncertain-query subset, and what are the VisDrone values of the reliability threshold and density-loss weight? Does the gain persist against a size-matched ordinary-attention or confidence-selected refinement control? Single targeted comparison is sufficient and a large new benchmark suite is not requested.
  3. Evaluation clarification: For AI-TOD-V2, please specify the backbone, feature levels, training schedule, and query settings of your method and the reimplemented baselines, and identify which differences remain for the quoted Dome-DETR results. It determines how much weights to give the 1.7 AP improvement.
5 comments

Thank you for your constructive feedback and for acknowledging the strengths of our work. We have carefully addressed the concerns raised and sincerely hope our responses clarify the contributions and improvements made. We would greatly appreciate your consideration of a higher score if you find our responses satisfactory. Thank you!

Weakness 1: The allocation comparison does not fully isolate the proposed mechanism.

A: Thank you for this valuable suggestion. To further investigate the contribution of the proposed allocation mechanism, we conducted additional controlled experiments, comparing different query allocation strategies under consistent decoder and reasoner configurations.

(1) We compared Global Top-K, Density-only, and Density + Scene Memory using 300 base queries and up to 500 supplementary queries, keeping the decoder and reasoner identical.

(2) As shown in Table 1, Density-only achieves 33.0 AP with 672 queries on average, outperforming Global Top-K (32.7 AP). Adding Scene Memory further improves AP to 33.5 with only 12 additional queries on average.

Table 1. Comparison of different query allocation strategies on VisDrone.

Query AllocationMean Total QP95 Total QMax Total QAP50AP
Global Top-K80080080053.532.7
Density-only67278880053.833.0
Density + Scene Memory68479280054.233.5

Weakness 2: The specific value of uncertainty-guided reasoning remains insufficiently separated from extra attention capacity.

A: Thank you for highlighting this concern. To further evaluate the contributions of uncertainty based query selection and our reasoning mechanism, we conducted additional controlled experiments, comparing our method with confidence based selection and ordinary attention under matched query settings.

(1) We compared random, confidence-based, and entropy-based query selection with identical selection sizes.

(2) We replaced the semantic-relational reasoning module with ordinary attention under comparable query and computational configurations. As shown in Table 2, our method achieves 33.5 AP, outperforming ordinary attention and confidence-based selection (33.0 AP).

(3) We also introduce ECE alongside NLL to evaluate calibration. Our method achieves the lowest NLL (1.14) and ECE (0.078), further supporting the effectiveness of the proposed reasoning mechanism.

Table 2. Controlled analysis of uncertainty-aware reasoning on VisDrone.

Reasoning StrategyQuery SelectionAPAP50Accu.NLLECE
Random SelectionRandom32.853.347.21.230.105
Confidence-basedLowest confidence33.053.648.11.200.096
Ordinary AttentionHighest entropy33.053.548.01.210.099
Proposed ReasonerHighest entropy33.554.250.21.140.078

Question 1: Controlled Comparison

A: Thank you for your valuable suggestion. Following your suggestion, we conducted additional controlled experiments with identical decoder and reasoner configurations. As shown in Table 1, Density + Scene Memory achieves 33.5 AP with 684 queries on average, outperforming Density-only (33.0 AP) by 0.5 AP and fixed-800-query Global Top-K (32.7 AP) by 0.8 AP. Together with the improved coverage in Table 7, these results support the complementary contribution of Scene Memory.

Question 2: Mechanism Clarification

A: Thank you for your valuable suggestions. We further clarify the implementation details of our reasoning module.

(1) The top 25% highest-entropy supplementary queries are selected according to Eq. (8), i.e., Ku = ceil(0.25 × Ks), where Ks denotes the number of supplementary queries.

(2) For VisDrone, the reliability threshold is 0.45 and the density-loss weight is 1.0.

(3) As shown in Table 2, our method outperforms confidence-based selection and ordinary attention in AP, NLL, and ECE, supporting the effectiveness of the proposed mechanism.

Question 3: Evaluation Clarification

A: Thank you for your valuable suggestion. We further clarify the experimental configurations on AI-TOD-V2.

(1) Our method uses COCO-pretrained RT-DETR-L, three feature levels (S3, S4, S5), 800 × 800 inputs, 300 base queries, and up to 500 supplementary queries.

(2) We specify the backbone, feature levels, query settings, and training schedules of the reimplemented baselines. All use the same dataset split, input resolution, and COCO-pretrained initialization. (3) Dome-DETR-L uses HGNetv2-B4 with four feature levels, whereas our method uses RT-DETR-L with three. Our method achieves 36.3 AP, exceeding the reported Dome-DETR result (34.6 AP) by 1.7 AP.

Dear Reviewer ToJY,

Thank you again for your valuable feedback. Following our previous responses in the rebuttal, we have further revised the manuscript, incorporating additional controlled experiments and clarifications of key implementation details. We sincerely hope these revisions address your concerns and further clarify our contributions. The corresponding changes are highlighted in blue.

Your responses partially address my main concerns. As a result, I am raising my rating from 2 to 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.