acceptodds
Under review as a conference paper at ICLR 2027

SEDEIM: Small-Object Enhancement in General Object Detection

Abstract

Detecting small objects in general object detection remains challenging in crowded scenes and near larger instances, where weak instance-specific cues can be obscured by dominant surrounding responses. To investigate this issue, we analyze DEIMv2 and find that missed small objects can still contain query-selected feature locations, indicating that spatial coverage alone is insufficient for successful detection. This motivates SEDEIM, a small-object-enhanced detector including the following three key components. Spatial-Enhanced Adapter supplements vision-transformer features with CNN details and spatially modulates their contribution. Context Decoupled Pyramid provides bidirectional cross-scale interaction while preserving a separate local-detail path to retain distinctions from neighboring instances. Local Evidence Refinement reuses features sampled by deformable attention to construct a localization-oriented representation for box prediction. Results on COCO val2017 show that SEDEIM-m achieves 54.2 AP and 35.5 APs, outperforming SOTA methods by 1.2 AP and 1.3 APs

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer Vision, Object Detection, Small object

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 2 reviews, above 85% of the others.
Reviewer_4DueOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents SEDEIM, an extension of DEIMv2 aimed at improving small-object detection in general scenes. The authors observe that many missed small objects are already covered by selected queries, suggesting that query coverage is not the only bottleneck. They introduce three modules to improve spatial detail extraction, multi-scale feature fusion, and localization refinement.

Critical strengths
  1. The paper identifies a meaningful failure mode beyond query coverage.

  2. Consistent improvements across model scales provide credible empirical evidence.

Critical weaknesses
  1. whether the added architectural complexity is justified by the improvements. For example, SEDEIM-M increases FLOPs from 52.3G to 67.9G while improving AP from 53.0 to 54.2. This is a meaningful gain, but without a compute-matched baseline, it is difficult to tell how much comes from the proposed mechanisms rather than additional processing capacity.

  2. limited novelty with weak evidence for the core causal claim. The three components mostly build on existing techniques (CNN spatial adaptation, bidirectional fusion, attention refinement); the contribution is an effective architectural combination rather than a distinct new method.

Questions for authors
  1. Would SEDEIM still outperform a simpler DEIMv2 variant with comparable FLOPs?

  2. Are the improvements particularly pronounced for small objects surrounded by larger or densely packed instances, as suggested by the motivation?

5 comments

We thank the reviewer for recognizing the strengths of our work, identifying its weaknesses, and providing constructive suggestions for improvement.

W1 & Q1 (Comparable Parameter Analysis)

Under comparable parameter and computational budgets, SEDEIM achieves clear gains over capacity-expanded DEIMv2-S baselines, demonstrating the benefits of our proposed mechanisms beyond simply increasing model capacity. Specifically, we evaluate straightforward capacity-expanded DEIMv2-S baselines by widening the original adapter CNN, the existing encoder fusion blocks, and both. Results are shown below

These capacity-expanded variants achieve 51.1, 51.2, and 51.3 AP, respectively, compared with 50.9 AP for DEIMv2-S. Their small-object AP values are 31.8, 31.4, and 31.7, whereas using our SEA and CDP achieves 51.7 AP and 32.0 APs and whole SEDEIM-S achieves 51.8 AP and 33.1 APs. Thus, ordinary capacity expansion recovers part of the improvement, but SEDEIM retains an advantage of 0.5 AP over the strongest overall capacity baseline and 1.4 APs over the strongest small-object capacity baseline.

ModelParams (M)FLOPs (G)APAPsAPmAPl
Baseline9.725.750.931.355.370.1
Baseline + wider adapter9.825.951.131.855.369.6
Baseline + SEA9.826.051.432.755.370.3
Baseline + wider encoder10.328.451.231.455.370.2
Baseline + CDP10.230.551.432.555.670.4
Baseline + wider adapter + encoder10.428.851.331.755.770.1
Baseline + SEA + CDP10.330.951.732.056.070.1
SEDEIM-S (Baseline + SEA + CDP + LER)10.330.951.833.156.070.4

W2 (Novelty)

SEDEIM introduces a problem-driven, coordinated design for preserving small-object distinctions and translating them into accurate localization. Its motivation is the gap between spatially covering a small object and obtaining an instance-specific representation that supports its detection.

The novelty is embodied in three specific information-processing mechanisms. SEA spatially modulates the CNN detail contribution using both semantic and detail responses. CDP extracts a separately projected local-contrast correction before mixing it with incoming cross-scale context. LER constructs a localization residual from the difference between localization-oriented and native scale allocations. Together, these mechanisms address how small-object information is introduced, preserved, and used for localization across successive detector stages.

The capacity-controlled experiments distinguish these mechanisms from straightforward architectural expansion. With the same reported 9.8M parameters and nearly identical FLOPs, SEA exceeds a wider-adapter baseline by 0.3 AP and 0.9 APs. The complete SEDEIM-S exceeds the jointly widened adapter-and-encoder baseline by 0.5 AP and 1.4 APs while using slightly fewer parameters. These comparisons establish benefits beyond simply increasing parameter capacity.

The interaction experiments (W2 & Q2 replied to Reviewer pGBb) further identify the functional contribution of the integrated design. Adding LER to SEA+CDP increases APs from 32.0 to 33.1 and AP75s from 34.1 to 35.4 while maintaining APm at 56.0. Controlled residual interventions additionally reveal that the effects of SAF modulation and CSDP local correction depend on one another. These results support complementary representation enhancement and localization refinement rather than an assumption of independently additive gains.

Independent detection results substantiate the targeted crowded-scene benefit. Under GT-defined neighborhoods, APs gains increase from 0.4 points for isolated small objects to 1.9 points for those with at least four neighbors. The crowded-small-object and medium/large-neighbor subgroups gain 2.0 and 1.7 APs, respectively. These neighborhood-stratified detection results complement the feature analyses and provide direct evidence that SEDEIM benefits the instance configurations motivating its design.

W2 & Q2 (Crowded-scene Verification):

SEDEIM delivers larger detection gains for small objects with more neighboring instances, with AP gains increasing from 0.4 points for isolated objects to 1.9 points for those with at least four neighbors.

We report APs and AP75s stratified by neighborhood instance count.

Specifically, following the spatial neighborhood defined in Appendix E.2, we expand each target box by a factor of 2.5 in both dimensions about its center and exclude the target box itself. Other valid annotated instance boxes with positive-area intersection with this ring are counted as neighbors, regardless of category or size. We partition all 15,264 valid small objects into groups with 0, 1, 2–3, and at least 4 neighbors and evaluate detections on the complete COCO val2017 set. Group membership depends exclusively on ground-truth geometry, without using feature measurements or detection outcomes.

SEDEIM-S improves AP over DEIMv2-S by 0.4, 0.8, 1.3, and 1.9 points, respectively, across the four neighborhood-count groups. The corresponding AP75 gains are 0.2, 0.8, 1.3, and 3.0 points. Aggregating objects with at least one neighbor yields a gain of 1.9 AP, compared with 0.4 AP for objects without neighbors in the ring. The subgroups with at least two small-object neighbors and with larger-instance neighbors improve by 2.0 and 1.7 AP, respectively. These results provide detection-level evidence, independent of our custom feature diagnostics, that the gains are more pronounced in neighboring and densely populated regions. They support our crowded-scene motivation without relying solely on improvements in the analysis metrics.

Neighborhood GroupSmall ObjectsDEIMv2-S APsSEDEIM-S APsDEIMv2-S AP75sSEDEIM-S AP75s
0 neighbors2,13719.019.4 (+ 0.4)20.821.0 (+ 0.2)
1 neighbor3,51425.526.3 (+ 0.8)28.128.9 (+ 0.8)
2-3 neighbors5,38720.421.7 (+ 1.3)21.622.9 (+ 1.3)
≥4 neighbors4,22622.124.0 (+ 1.9)21.824.8 (+ 3.0)
≥1 neighbor13,12730.432.3 (+ 1.9)32.534.9 (+ 2.4)
≥2 small-object neighbors5,63220.422.4 (+ 2.0)21.223.0 (+ 1.8)
≥1 medium/large-object neighbor8,95128.430.1 (+ 1.7)29.932.2 (+ 2.3)

Dear Reviewer 4Due,

Thank you again for your detailed review. As the discussion phase ends on November 18 (AoE), we would like to confirm whether our responses have addressed your concerns. In summary, our responses include comparable parameter analysis, discussion and clarity of our novelty, and crowded-scene verification.

We are happy to discuss any remaining questions. If our responses and the revised paper have addressed your concerns, we would be grateful if you would consider updating your evaluation.

Sincerely, The Authors

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

The paper studies why small objects fail in general-scene, real-time DETR-style detection and proposes SEDEIM on top of DEIMv2 (DINOv3 ViT backbone, Efficient Hybrid Encoder, deformable-attention decoder). A diagnostic analysis on COCO shows that 70.2% of missed small objects already contain query-selected tokens, that missed objects are less separable from their neighborhoods and become less separable after encoding, that representations respond to nearby content almost as strongly as to the object itself, and that tracked candidates of missed objects barely improve their box IoU during decoding. Three modules follow from this: (1) SEA, a CNN detail trunk added to the Spatial Tuning Adapter, with a spatially adaptive fusion (SAF) that modulates the detail branch at stride 8 via a learned single-channel modulation map; (2) CDP, a four-stage bidirectional cross-scale pyramid after the encoder whose CSDP blocks add a neighbor-contrast (local mean subtraction) branch to preserve instance distinctions; (3) LER, which reuses existing multi-scale deformable-attention samples and weights to compute a localization-oriented reweighted aggregation, applied as a bounded residual only to the box-regression features. On COCO val2017, SEDEIM improves over DEIMv2 at all four scales, e.g., +1.2 AP / +1.3 AP_S at M scale and +0.9 AP / +1.8 AP_S at S scale, while AP_M and AP_L also increase.

Critical strengths
  1. The motivating analysis goes beyond the usual coverage argument and is well documented: the four diagnostics (query-selection coverage, target-neighborhood separation, matched perturbation sensitivity, candidate IoU tracking) have explicit protocols in Appendix E and genuinely identify where the pipeline loses small objects.
  2. The experimental study is thorough for this paper type: four model scales with consistent gains, component-wise and pairwise ablations, SAF scope selection, CDP interaction-order ablation, LER design-choice ablation, and feature-map visualizations that match the diagnostics.
  3. The gains are meaningful and not small-object-specific: AP_S rises by 1.3-1.8 points across scales while AP_M and AP_L also improve, indicating the changes do not trade larger objects away.
Critical weaknesses
  1. Efficiency is not measured. FLOPs grow by roughly 20-30% at every scale (S: 25.7→30.9G; M: 52.3→67.9G; X: 151.7→183.2G), yet the paper positions SEDEIM within real-time detection and reports no latency or FPS; at M scale the +15.6G FLOPs for +1.2 AP needs a runtime justification.
  2. The module interactions are not clean: in Table 2, pairwise SEA+CDP yields 32.0 AP_S, below either module alone (32.7 and 32.5), and only adding LER restores 33.1. The text describes the three modules as improving different stages, but this interference between the two representation modules is not analyzed or explained.
  3. The analysis and the design validation share the same metrics. The separation and perturbation-sensitivity measures were constructed together with the modules, so improvements on them are partly built in; the crowded-scene motivation is never tested with an independent measure such as AP stratified by neighborhood instance density, although the paper's own Appendix E.2 defines exactly such neighborhoods.
  4. Evaluation rests on COCO val2017 only. For a paper claiming improved small-object detection in general scenes, a second benchmark (e.g., VisDrone or AI-TOD for scale transfer, or a crowded-scene set) and a comparison with the general-scene small-object method TinyFormer are absent.
Questions for authors
  1. Report latency/FPS for SEDEIM and DEIMv2 at each scale under the same protocol. Is the accuracy-overhead trade-off still favorable at M and X scales?
  2. Explain the SEA+CDP interference in Table 2 (pairwise AP_S below either single module). Does the order or position of CDP relative to the encoder interplay with SAF at F3?
  3. Using the neighborhood definition already in Appendix E.2, report the detection gains stratified by the number or proximity of surrounding instances, to verify the crowded-scene claim quantitatively.
  4. LER alone adds 0.1 AP and its variants in Table 6 reduce AP below the baseline while slightly lowering AP_M. Is this machinery justified, and does the box-only correction hurt medium objects in the full model?
6 comments

We thank the reviewer for recognizing the strengths of our work, identifying its weaknesses, and providing constructive suggestions for improvement.

We first present the revised table of the main results, including corrections to the results for all scales of YOLO26 (pointed out by Area Chair 1DTq) and the latency of each model. Moreover, we add the results of TinyFormer. All scales of YOLO26 are first pretrained on Objects365V1, then fine-tuned on COCO. We measure both accuracy and latency of YOLO26 with NMS.

ModelParams (M)FLOPs (G)Latency (ms)APAPsAP50AP75APmAPl
YOLO26-S9.520.91.5848.629.565.852.853.265.9
DEIMv2-S9.725.72.7350.931.468.355.155.370.3
TinyFormer-S10.832.62.8551.532.668.955.755.369.8
SEDEIM-S10.330.93.2251.833.168.756.256.070.4
YOLO26-M20.468.43.2053.136.870.757.757.768.9
DEIMv2-M18.152.33.2753.034.270.257.657.471.5
TinyFormer-M20.263.93.3853.735.170.958.357.871.3
SEDEIM-M20.267.93.8954.235.571.159.158.272.4
YOLO26-L24.886.83.5755.138.572.660.159.671.1
DEIMv2-L32.296.83.9456.037.673.561.160.974.9
TinyFormer-L33.6105.94.0656.839.074.361.861.275.5
SEDEIM-L34.8113.94.6156.939.074.362.061.675.2
YOLO26-X55.7194.44.4557.541.875.062.862.173.2
DEIMv2-X50.3151.74.7357.839.175.363.262.975.9
TinyFormer-X51.5164.24.9158.540.975.964.163.276.6
SEDEIM-X55.5183.25.5058.439.975.764.063.476.3

W1 & Q1 (Efficiency Measurement)

We report latency for DEIMv2 and SEDEIM at all scales under the same inference protocol in the table below. SEDEIM offers an accuracy–speed trade-off for general object detection, particularly small objects, while retaining real-time inference on the evaluated hardware. Notably, SEDEIM-S improves AP by 0.9 and APs by 1.8 with only 0.49 ms of additional latency.

ScaleDEIMv2 (ms)SEDEIM (ms)Increased latency
S2.733.22+0.49 ms (17.9%)
M3.273.89+0.62 ms (19.0%)
L3.944.61+0.67 ms (17.0%)
X4.735.50+0.77 ms (16.3%)

W2 & Q2 (Module Interactions Analysis)

We attribute the reduced small-object gains of SEA+CDP to the loss of fine spatial information during cross-scale fusion of SEA-enhanced features. SEA supplies additional spatial details, while CDP integrates information across resolutions. Their successive updates introduce a trade-off between contextual integration and the spatial precision required to localize small instances, which is also reflected in the balance between small-object accuracy and performance on medium and large objects.

We additionally evaluate small-object detection at IoU thresholds of 0.50 and 0.75, denoted as AP50s and AP75s, respectively, to examine the module interactions under different localization requirements. The results are shown below. Both modules individually outperform the baseline, and their combination also remains above it: SEA+CDP achieves 32.0 APs versus 31.4 for the baseline. The interaction is more pronounced under stricter localization requirements. Relative to SEA alone, the combination decreases AP50s by 0.3 but AP75s by 0.9; relative to CDP alone, the corresponding decreases are 0.2 and 0.4. This pattern supports our interpretation that the combined representation updates particularly affect precise small-object localization.

LER complements these representation updates with a localization-oriented residual constructed from features already sampled by deformable attention. Adding LER raises APs from 32.0 to 33.1, AP50s from 50.2 to 50.6, and AP75s from 34.1 to 35.4. The larger recovery at AP75s highlights its contribution to precise localization. The full model exceeds the baseline by 1.7 APs and 1.7 AP75s, demonstrating the benefit of combining representation enhancement with dedicated localization refinement.

ModelAPsAP50sAP75s
Baseline31.449.933.7
+ SEA32.750.535.0
+ CDP32.550.434.5
+ SEA + CDP32.050.234.1
+ SEA + CDP + LER33.150.635.4

Q2 (CDP Position):

We evaluate CDP before and after the Hybrid Encoder, both with and without SEA, in the table below. SAF and CDP operate along a connected feature pathway: SAF modulates the F3 representation that is subsequently processed by the encoder and CDP. We therefore examine whether changing CDP’s placement alters the interaction. Post-encoder placement consistently achieves better overall and small-object accuracy. Compared with pre-encoder placement, it improves AP/APs by 0.3/1.3 points for CDP alone and by 0.4/1.0 points for SEA+CDP. Adding SEA increases overall AP at both placements, while APs decreases from 32.5 to 32.0 after the encoder and from 31.2 to 31.0 before it. Thus, relocating CDP does not eliminate the small-object trade-off between the representation modules. These results support the proposed post-encoder placement and show that the interaction is not specific to placing CDP after the encoder.

ConfigurationCDP PlacementAPAPsAPmAPl
Baseline + CDPAfter the Hybrid Encoder51.432.555.670.4
Baseline + CDPBefore the Hybrid Encoder51.131.255.369.9
Baseline + SEA + CDPAfter the Hybrid Encoder51.732.056.070.1
Baseline + SEA + CDPBefore the Hybrid Encoder51.331.055.670.0

W3 & Q3 (Crowded-scene Verification):

SEDEIM delivers larger detection gains for small objects with more neighboring instances, with AP gains increasing from 0.4 points for isolated objects to 1.9 points for those with at least four neighbors.

We report APs and AP75s stratified by neighborhood instance count.

Specifically, following the spatial neighborhood defined in Appendix E.2, we expand each target box by a factor of 2.5 in both dimensions about its center and exclude the target box itself. Other valid annotated instance boxes with positive-area intersection with this ring are counted as neighbors, regardless of category or size. We partition all 15,264 valid small objects into groups with 0, 1, 2–3, and at least 4 neighbors and evaluate detections on the complete COCO val2017 set. Group membership depends exclusively on ground-truth geometry, without using feature measurements or detection outcomes.

SEDEIM-S improves AP over DEIMv2-S by 0.4, 0.8, 1.3, and 1.9 points, respectively, across the four neighborhood-count groups. The corresponding AP75 gains are 0.2, 0.8, 1.3, and 3.0 points. Aggregating objects with at least one neighbor yields a gain of 1.9 AP, compared with 0.4 AP for objects without neighbors in the ring. The subgroups with at least two small-object neighbors and with larger-instance neighbors improve by 2.0 and 1.7 AP, respectively. These results provide detection-level evidence, independent of our custom feature diagnostics, that the gains are more pronounced in neighboring and densely populated regions. They support our crowded-scene motivation without relying solely on improvements in the analysis metrics.

Neighborhood GroupSmall ObjectsDEIMv2-S APsSEDEIM-S APsDEIMv2-S AP75sSEDEIM-S AP75s
0 neighbors2,13719.019.4 (+ 0.4)20.821.0 (+ 0.2)
1 neighbor3,51425.526.3 (+ 0.8)28.128.9 (+ 0.8)
2-3 neighbors5,38720.421.7 (+ 1.3)21.622.9 (+ 1.3)
≥4 neighbors4,22622.124.0 (+ 1.9)21.824.8 (+ 3.0)
≥1 neighbor13,12730.432.3 (+ 1.9)32.534.9 (+ 2.4)
≥2 small-object neighbors5,63220.422.4 (+ 2.0)21.223.0 (+ 1.8)
≥1 medium/large-object neighbor8,95128.430.1 (+ 1.7)29.932.2 (+ 2.3)

W4 (VisDrone and TinyFormer results):

We provide complete comparisons with TinyFormer across the S, M, L, and X scales on both COCO (the first table) and VisDrone (table below). For VisDrone, we follow the same COCO-pretrained fine-tuning strategy as TinyFormer, using the corresponding COCO-trained checkpoints and the same training strategy to provide a controlled comparison.

On COCO, SEDEIM achieves stronger AP and APs at the S and M scales while enhancing APm and APl. SEDEIM-S exceeds TinyFormer-S by 0.3 AP and 0.5 APs, while SEDEIM-M achieves margins of 0.5 AP and 0.4 APs. APm reaches 56.0/58.2 for S/M, compared with TinyFormer’s 55.3/57.8, and APl reaches 70.4/72.4 versus 69.8/71.3. Overall AP remains comparable at L/X.

The VisDrone results further support SEDEIM’s cross-dataset transferability. Following the same COCO-pretrained fine-tuning strategy as TinyFormer, SEDEIM-S/M achieve 29.2/31.2 AP, compared with TinyFormer’s 28.9/30.9 AP, with corresponding AP50 margins of 0.5/0.7 points. Higher AP and AP50 are also observed at L/X. These results indicate that the design transfers effectively from general-scene detection to aerial scenes containing numerous small and crowded objects.

A key architectural strength of SEDEIM is its coordinated treatment of spatial detail, local distinctions, and instance localization. TinyFormer emphasizes additional high-resolution features through its Spatial Semantic Adapter and parallel multi-scale fusion through PBM. SEDEIM instead combines spatially adaptive detail modulation at F3, a post-encoder Context Decoupled Pyramid that preserves local distinctions during cross-scale interaction, and Local Evidence Refinement that reuses existing deformable-attention samples for localization supplementation. Together, these mechanisms address how small-object information is both retained and used for accurate box prediction.

ModelParams (M)FLOPs (G)APAP50
DEIMv2-S9.725.726.745.0
TinyFormer-S10.832.628.947.9
SEDEIM-S10.330.929.248.4
DEIMv2-M18.152.328.948.0
TinyFormer-M20.263.930.950.0
SEDEIM-M20.267.931.250.7
DEIMv2-L32.296.830.449.9
TinyFormer-L33.6105.932.252.1
SEDEIM-L34.8113.932.452.5
DEIMv2-X50.3151.732.252.3
TinyFormer-X51.5164.234.755.5
SEDEIM-X55.5183.234.955.8

Q4 (LER explanation):

LER is justified by its contribution to precise small-object localization. Although its standalone overall AP gain is 0.1, it improves APs from 31.3 to 32.0. Table 6 supports the complete design: removing size-dependent scaling or native context subtraction reduces APs to 31.5 or 31.9, respectively, and both variants reduce overall AP from 51.0 to 50.8. More importantly, adding LER to SEA+CDP increases APs from 32.0 to 33.1 and AP75s from 34.1 to 35.4, demonstrating its complementary localization benefit within the complete detector.

The box-only correction does not reduce medium-object accuracy in the full model. The small standalone APm decrease from 55.3 to 55.1 does not persist in the combined configurations: adding LER to SEA increases APm from 55.3 to 55.9, and adding it to CDP increases APm from 55.6 to 55.7. In the complete model, adding LER maintains APm at 56.0 while improving small-object localization.

Comments on the paper

1 comment

We thank the Area Chair and reviewers for their constructive feedback. We summarize below the additional experiments, the questions they address, and the corrections and clarifications provided during rebuttal.

1. Detection-level verification of the crowded-scene motivation. We added neighborhood-stratified APs, AP75s, and ARs using groups defined exclusively by ground-truth geometry, independently of our feature-analysis metrics. APs gains increase from 0.4 points for isolated small objects to 1.9 points for those with at least four neighbors, where AP75s and ARs also increase by 3.0 and 4.1 points. Small objects with densely packed small neighbors or medium/large neighbors gain 2.0 and 1.7 APs, respectively. These results directly substantiate the targeted benefit in challenging neighborhoods.

2. Benefits beyond additional model capacity. We evaluated DEIMv2-S controls with a wider adapter, wider encoder fusion blocks, and both. Compared with these parameter-comparable controls, SEDEIM-S retains advantages of 0.5 AP and 1.3 APs over the strongest respective controls. Together with the component ablations, these comparisons support the benefits of the proposed mechanisms beyond straightforward capacity expansion.

3. Novelty of the coordinated design. We clarified that SEDEIM addresses the gap between spatially covering small objects and obtaining representations that support their accurate localization. SEA introduces semantic-conditioned spatial detail modulation; CDP preserves separately projected local contrasts alongside cross-scale contextual fusion; and LER constructs a localization residual from the difference between localization-oriented and native scale allocations. These mechanisms form a coordinated system for introducing, preserving, and using small-object information, rather than simply combining standard architectural blocks. The additional capacity controls, interaction analyses, and independent neighborhood-stratified detection results provide experimental support for this problem-driven design.

4. Runtime and memory measurements. We added TensorRT FP16 and PyTorch FP16 latency, together with PyTorch peak allocated memory, for DEIMv2 and SEDEIM at all four scales. Measurements use the same A100-40GB GPU, 640 × 640 inputs, and batch size 1. SEDEIM adds 0.49–0.77 ms of TensorRT latency, corresponding to 16–19% overhead. These measurements explicitly quantify the accuracy–speed trade-off, with particularly favorable accuracy gains at S/M.

5. TinyFormer comparisons and an additional benchmark. We added complete S/M/L/X comparisons on COCO and VisDrone. On COCO, SEDEIM exceeds TinyFormer by 0.3/0.5 AP and 0.5/0.4 APs at S/M, while also attaining higher APm and APl; overall AP is comparable at L/X. Following the same COCO-pretrained fine-tuning strategy, SEDEIM exceeds TinyFormer on VisDrone at every scale, supporting its transferability to aerial scenes with numerous small objects.

6. Module interactions and CDP placement. Additional AP50s/AP75s evaluation shows that the SEA+CDP deficit is larger under stricter localization requirements. Adding LER increases APs from 32.0 to 33.1 and AP75s from 34.1 to 35.4 while maintaining APm at 56.0, supporting its complementary localization role. We also evaluated CDP before and after the Hybrid Encoder, with and without SEA. Post-encoder placement consistently performs better, while relocating CDP does not eliminate the non-additive small-object behavior. These experiments clarify the interactions within the coordinated representation-enhancement and localization-refinement framework.

7. Corrections to reported results. We corrected the COCO-to-YOLO annotation conversion and reevaluated all four YOLO26 checkpoints. We also corrected baseline APs to 31.4, LER-only APm to 55.1 using the intended checkpoint, and the SEA-without-SAF gain to 1.0 APs. The complete SEDEIM-S improvement is 1.7 APs, and the affected comparisons have been updated consistently.

8. Additional implementation and analysis details. We provided single-cue LER inference experiments and scale-reference sensitivity tests on both COCO and VisDrone. Combining all four cues achieves the best APs among the tested LER configurations, while overall AP varies by only 0.1 points across a 16-fold scale-reference range on each dataset. We further clarified the effective entropy coefficient, cosine-similarity computation, SAF initialization and bounds, the Figure 1(c) perturbation protocol, and the final query assignments underlying spatially covered missed objects.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.