acceptodds
Under review as a conference paper at ICLR 2027

LiG-DETR: Local-in-Global Reassembly in Latent Space for Aerial Object Detection

Abstract

Aerial object detection faces substantial scale and density variations. Small objects are easily degraded by downsampling and feature compression, while medium and large objects require sufficient global context. Existing methods mainly follow two paradigms: image slicing provides clearer local evidence but relies on independent crop-level prediction and post-processing, whereas feature- and query-level optimization preserves unified inference but operates on already compressed full-image representations, limiting recovery of fine-grained information. This raises a key question: can aerial detection directly acquire high-fidelity local evidence before feature degradation and integrate it into a unified end-to-end framework? To this end, we propose LiG-DETR, an Efficient Global-Local Reassembly framework that reformulates image slicing as high-fidelity local feature acquisition. A shared encoder extracts global and locally magnified features, which are projected into the detector feature space. The projected local features are reassembled according to their original spatial locations to form a globally aligned local feature level, and a single DETR decoder jointly decodes global and local features. To reduce redundant computation, Context-Preserved Selective Reassembly focuses high-resolution encoding on informative regions while preserving a dense feature layout, and Density-Aware Adaptive Query Allocation adapts the decoder query budget using encoder proposal scores. Experiments show substantial gains on small and medium objects while retaining strong large-object performance, with favorable accuracy–efficiency trade-offs and improved cross-domain generalization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Aerial Object Detection, Global-Local Reassembly, Small Object

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_oA5DOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to allocate computation more selectively without sacrificing detection performance. Specifically, CPSR uses Laplacian-based high-frequency scores to select informative regions for local zoom-in re-encoding and then reassembles them with global features to preserve contextual information. DA-AQA further adapts the number of decoder queries according to the confidence distribution of encoder proposals, allowing different images to use different decoding budgets based on their estimated object density.

Critical strengths
  • The motivation is clear and well aligned with practical challenges in remote-sensing object detection, where high-resolution imagery, small objects, and large variations in object density make computational efficiency an important problem.
  • The experimental section compares against a relatively broad set of recent and competitive baselines.
Critical weaknesses
  • The design of Context-Preserved Selective Reassembly (CPSR) is not sufficiently justified, and several important implementation choices remain unclear or heuristic. Specific concerns include:

    1. The relationship among the 2×22\times2 local crops introduced in the full local reassembly scheme, the N×NN\times N fine-grained patches used in CPSR, and the subsequent G×GG\times G packing operation is not clearly explained. These three partitioning schemes serve different purposes, but the current presentation makes their roles, magnification factors, and spatial correspondence difficult to follow. The authors should reorganize this part and explicitly clarify how these operations are related and how the packed observations are mapped back to the original image coordinate system.
    2. The use of a Laplacian-based high-frequency score for patch selection appears highly heuristic. High-frequency response does not necessarily correspond to object-rich or informative regions, particularly in remote-sensing imagery. The paper should provide direct evidence that the proposed score indeed prioritizes useful regions, e.g., by reporting GT coverage or small-object recall under different selection budgets.
    3. Packing spatially non-adjacent patches into the same local observation may introduce artificial spatial context. Since DINOv2 relies on global self-attention and positional embeddings, patches that are distant in the original image can interact after being packed next to each other, creating dependencies that do not exist in the original scene. Such effects cannot be removed simply by reassembling the encoded features back to their original locations. The paper should evaluate whether this artificial interaction affects representation quality, for example by comparing packed encoding with independent patch encoding or with full local reassembly.
    4. The direct reassembly of local and global features is insufficiently analyzed. CPSR replaces selected regions with zoomed-in local features while filling unselected regions with G0G_0, although these two feature sources are produced under different spatial scales and contextual conditions and may therefore exhibit distribution mismatch or boundary discontinuities. Moreover, the final feature set still uses {G0,G1,L}\{G_0,G_1,L\}, while a large portion of LL is identical to G0G_0, potentially introducing redundant computation. Additional ablations are needed to justify both the hard replacement strategy and the necessity of retaining G0G_0 and LL simultaneously.
  • The backbone configurations used in the experiments are not sufficiently specified, which raises concerns about the fairness of the comparisons. The implementation details only state that LiG-DETR uses a pretrained DINOv2 visual encoder, but the specific DINOv2 variant (e.g., ViT-S/14, ViT-B/14, or ViT-L/14) is not reported. More importantly, it is unclear whether all methods compared in Tables 1 and 2 use the same visual backbone and comparable model capacity. If different backbone variants are used across methods, the reported performance gains may partially originate from stronger pretrained representations rather than the proposed CPSR/DA-AQA modules. The authors should explicitly report the backbone, parameter count, and input resolution for every compared method, and ideally provide controlled comparisons under the same backbone and training setting to isolate the contribution of the proposed method.

Questions for authors

See Weakness

8 comments

We thank the reviewer for the careful reading and constructive suggestions. We have revised the manuscript accordingly.

R1. Local Patch Selection, Packing, and Feature Reassembly

Given a full-image input of size H×WH\times W, the global branch encodes the full image, and the local branch proceeds in three steps.

  1. Partitioning and selection. We divide the full image into an N×NN\times N grid of patches, select patches based on their scores and the budget, and record their positions in the original grid. For example, when N=4N=4, the full image is divided into 16 patches.
  2. Packing and upscaling. We sort the selected patches by score and pack each group of G2G^2 patches into a local image with a G×GG\times G layout. Geometrically, each packed local image is upscaled from (GH/N)×(GW/N)(GH/N)\times(GW/N) to the encoder input size of H×WH\times W. When N=4N=4 and G=2G=2, every four patches form one local image, which is enlarged by a factor of 2 along both height and width. Selecting 4, 8, 12, or 16 patches yields 1, 2, 3, or 4 local inputs, respectively. Selecting all 16 patches covers the full image.
  3. Encoding and reassembly. We pass the upscaled local inputs through the shared encoder and project the resulting features. We then place each patch's features in the corresponding region of LL using the position recorded in Step 1. For example, in Figure 2, patch 6 is packed into the upper-right slot, and its features are placed back in row 2, column 2 of the original grid. We fill unselected regions with the corresponding global features from G0G_0. A single decoder then jointly uses G0,G1,L\\{G_0,G_1,L\\} to detect objects across the full image.

If you find this explanation clearer, we will revise the final manuscript accordingly.

R2. Laplacian Patch Selection Prioritizes Small-Object Coverage

On the VisDrone-DET2019 validation set, selecting only 25% of the patches in each image using Laplacian scores covers 48.3% of small objects, 1.93 times the expected coverage of uniform random selection under the same budget. We evaluated 26,588 small objects across all validation images, counting an object as covered if the center of its ground-truth bounding box fell within a selected patch. Coverage reaches 79.9% and 94.0% when the selection budget increases to 50% and 75% of the patches, respectively. Across the three budgets, coverage exceeds the random-selection expectation by 23.3, 29.9, and 19.0 percentage points, respectively. These results show that, under the same local encoding budget, Laplacian scoring allows regions containing more small objects to receive high-resolution encoding.

Laplacian scoring is a replaceable patch selection strategy within the Local-in-Global Reassembly framework. The core contribution of this work is to reassemble high-resolution local features from the selected regions at their original image locations and jointly decode them with the global features.

R3. Laplacian Score-Based Packing Achieves Detection Accuracy Comparable to Spatially Adjacent Packing

On the VisDrone-DET2019 validation set, we use the same model weights and vary only the packing layout at inference. Both settings use the same 16 patches and encode four local inputs per image (B=4B=4), each upscaled by a factor of 2 along both height and width.

Packing methodLocal input constructionAP (%)APₛ (%)
Laplacian score-based packing (default)Pack patches in groups of four in score order and upscale each group37.1227.56
Spatially adjacent packingPack four neighboring patches in their original relative positions and upscale the group37.4427.85

Under the same encoding budget, the default score-based packing achieves 37.12 AP and 27.56 APₛ, differing from spatially adjacent packing by only 0.32 and 0.29 percentage points, respectively. These results support CPSR's ability to pack patches from different parts of the original image into a single local input while maintaining detection accuracy comparable to that of spatially adjacent packing.

R4. Global Feature Filling and the Complementarity of G0G_0 and LL

Joint optimization of cross-scale features. Zoomed-in local encoding and whole-image encoding operate at different scales and with different context, so the reassembled feature map LL may exhibit feature differences at region boundaries. To help the two feature streams work together, we use the same trainable visual encoder to extract both sets of features and align them with their original image locations after projection. When only a subset of patches is selected, the decoder receives LL composed of both types of features during training. Feature extraction and decoding are jointly optimized end to end under the same detection objective. This allows the decoder to learn how to combine cross-scale features on both sides of a boundary for detection.

Why we fill with G0G_0. The decoder samples from a regular two-dimensional feature map, so LL needs to retain the spatial grid of the full image. Because G0G_0 and LL are spatially aligned, we can directly reuse the whole-image features from the corresponding locations in G0G_0. This allows the decoder to access image information from unselected regions when sampling from LL. On the VisDrone-DET2019 validation set, we use the same B=1B=1 model weights and change only how unselected regions are filled at inference. Compared with zero filling, filling with G0G_0 increases AP from 22.9 to 34.6 and APₛ from 16.5 to 25.3, corresponding to gains of 11.7 and 8.8 percentage points, respectively. This result supports reusing existing whole-image features to fill unselected regions.

Why we retain both G0G_0 and LL. The two levels share features in unselected regions, and retaining both increases the amount of feature data processed by the decoder. In selected regions, LL provides local details from zoomed-in encoding, and the complete G0G_0 preserves the representation of the same locations in the context of the whole image. Joint decoding can therefore use both local details and global context at the same location. Table 6 shows that adding G0G_0 as a separate feature level to G1,L\\{G_1,L\\} increases AP from 32.9 to 35.5 and APₛ from 23.3 to 26.4, supporting the use of both feature levels.

R5. Backbone Configurations and Same-Backbone Comparison

On VisDrone-DET2019, LiG-DETR and RF-DETR-L both use a windowed DINOv2-Small backbone with a patch size of 16, with the same pretrained weights, 704×704704\times704 inputs, training data, number of training epochs, learning rate, and effective batch size. In this comparison, LiG-DETR improves AP from 29.4 to 37.1 (+7.7 percentage points) and small-object AP from 18.6 to 27.6 (+9.0 percentage points). Table 6 further shows that adding only the reassembled local feature LL raises AP from 29.4 to 32.9, directly quantifying the benefit of local feature reassembly with the same backbone.

The other compared methods retain their respective backbones and detection architectures. Table 2 of the revised manuscript lists the backbone and parameter count of each method compared on VisDrone, and Section 4.1 describes the input resolution settings. The new matched-resolution comparison shows that at 704×704704\times704, LiG-DETR achieves an overall AP of 37.1, compared with 36.0 for Dome-DETR-L, and leads by 3.4 percentage points in medium-object AP (49.9 vs. 46.5).

Thank you for the clarifications and additional experiments. They partially address my concerns about the experimental setup and comparison fairness. However, I believe the methodological novelty remains limited, as the contribution appears primarily to be an engineering integration of local cropping, feature reassembly, and adaptive query allocation. The gains over the original RF-DETR baseline entail substantially greater computation and lower inference throughput. I acknowledge the improved trade-off relative to the slicing baseline, but the incremental accuracy advantage remains modest and does not sufficiently strengthen the contribution in my assessment. I note the commitment to release code in the responses to other reviewers, but a future release promise cannot substitute for reproducibility materials available during review.

Reviewer_QHS2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes LiG-DETR, a global-local reassembly framework for aerial object detection. It extracts high-resolution local features and reassembles them into a unified global feature space for DETR-based prediction. The framework further introduces selective local encoding and adaptive query allocation to improve computational efficiency. Experiments on VisDrone-DET2019, AI-TOD-v2, and DGDet-VisDrone demonstrate substantial improvements, particularly for small-object detection.

Critical strengths
  1. The paper presents a practical approach to integrating high-resolution local evidence into a single DETR decoder without independent crop-level predictions.

  2. LiG-DETR achieves substantial accuracy gains on two aerial detection benchmarks and demonstrates improved performance under cross-domain evaluation.

Critical weaknesses
  1. The core idea of extracting high-resolution local features and combining them with global representations is closely related to existing slicing and multi-scale detection strategies. Although unified decoding is a useful design, the paper does not sufficiently establish its conceptual novelty beyond integrating these established techniques.

  2. CPSR assumes that high-frequency responses indicate regions requiring high-resolution encoding. However, high-frequency backgrounds may dominate the selection, while small low-contrast objects may be overlooked. The paper lacks direct evidence that this criterion reliably selects informative object regions.

  3. DA-AQA determines the query budget from cumulative proposal confidence, but confidence distributions do not necessarily reflect object density. Dense scenes with low-confidence objects or sparse scenes with many background proposals may lead to inappropriate query allocation. This assumption is not sufficiently validated.

  4. Although the paper provides extensive component ablations and a full-image upscaling comparison, it lacks a controlled comparison with independent crop-level detection under comparable computational budgets. Consequently, the advantages of latent reassembly over conventional slicing, particularly in accuracy-efficiency trade-offs, remain insufficiently established.

Questions for authors

see the Weaknesses

1 comment

R1. From High-Resolution Local Encoding to Unified Whole-Image Prediction

The core contribution of LiG-DETR is to combine zoomed-in local encoding and whole-image prediction in a single detection pipeline. Conventional slicing makes predictions independently on each image crop and then merges the detection boxes. Conventional multi-scale fusion exchanges information between previously extracted feature levels. LiG-DETR zooms in on selected image patches before encoding to obtain finer-grained local features, then reassembles them at their original image locations to form a feature level aligned with the whole-image features. This allows a single decoder to use both high-resolution local details and whole-image context in one whole-image prediction pass.

Table 2 of the revised manuscript directly compares these two ways of using local inputs. With the same backbone and input resolution and comparable computational cost, LiG-DETR outperforms the RF-DETR-L slicing baseline by 1.6 AP points (37.1 vs. 35.5) and achieves approximately 2.1 times its end-to-end throughput (15.6 vs. 7.5 FPS). These results show that reassembling local features before prediction improves whole-image detection accuracy and reduces the overhead of per-crop prediction and box merging.

R2. Laplacian Patch Selection Prioritizes Small-Object Coverage Under Limited Budgets

To test whether Laplacian scoring prioritizes regions containing small objects in real images, we evaluated 26,588 small objects in the VisDrone-DET2019 validation images. We counted an object as covered if the center of its ground-truth bounding box fell within a selected patch.

When only 25% of the patches in each image are selected, Laplacian patch selection covers 48.3% of small objects, which is 1.93 times the expected coverage of uniform random selection under the same budget. When 50% and 75% of the patches are selected, coverage reaches 79.9% and 94.0%, exceeding the expected coverage of random selection by 29.9 and 19.0 percentage points, respectively. This shows that, under the same local encoding budget, Laplacian scoring enables high-resolution encoding for regions covering more small objects. Unselected regions are still filled with global features from the corresponding locations, so they remain covered by whole-image detection.

R3. DA-AQA Query Allocation in Two Challenging Scenarios

To test whether these two challenging scenarios lead to inappropriate query allocation, we compare DA-AQA with a fixed budget of 500 queries per image on the AI-TOD-v2 test set using the same model weights.

Challenging scenario on AI-TOD-v2ImagesMean objects per imageAR₅₀ with 500 fixed queriesDA-AQA AR₅₀Mean DA-AQA queries per image
Dense, low-confidence scenes234249.552.35%66.78%887.5
Sparse scenes with many background proposals2,2122.1283.91%82.15%358.5

Dense, low-confidence scenes (234 images). DA-AQA allocates an average of 887.5 queries, increasing AR₅₀ from 52.35% to 66.78%. The additional queries improve recall by 14.43 percentage points.

Sparse scenes with many background proposals (2,212 images). DA-AQA uses an average of 358.5 queries, a 28.3% reduction from the fixed budget of 500. AR₅₀ decreases from 83.91% to 82.15%, a drop of 1.76 percentage points.

These results show that, in the two challenging scenarios evaluated, DA-AQA uses additional queries to improve recall in dense scenes and substantially reduces the query count in sparse scenes with many background proposals at a small cost in recall.

R4. Feature Reassembly Is More Accurate and Faster Than Independent Slicing at Comparable Computational Cost

To directly compare the two ways of using high-resolution local information, we added an RF-DETR-L slicing baseline on the VisDrone-DET2019 validation set (Table 2 of the revised manuscript). Both methods use the same backbone and encode the full image once and four local inputs per image, each at an input resolution of 704 × 704. The slicing baseline makes independent predictions on the full image and its four quadrants, then maps the detection boxes back to the original image and merges them using NMS. LiG-DETR reassembles the local information at the feature level and uses a single decoder to make whole-image predictions.

At comparable computational cost (645.7 vs. 627.8 GFLOPs), LiG-DETR improves AP from 35.5 for the slicing baseline to 37.1 (+1.6 percentage points) and small-object AP from 26.2 to 27.6 (+1.4 percentage points). On the same RTX 3090, LiG-DETR achieves 15.6 FPS, approximately 2.1 times the slicing baseline's 7.5 FPS. This comparison shows that, at comparable computational budgets, reassembling features before unified prediction lets the decoder combine local details with whole-image context to improve accuracy while eliminating the overhead of repeated decoding and box merging.

Comments on the paper

4 comments

We thank the reviewer for the careful reading and constructive suggestions. We have revised the manuscript in response to the feedback and address related comments together below.

R1. Framework Novelty

The core innovation of LiG-DETR is to reassemble high-resolution local features in the original image coordinate system before making unified whole-image predictions. Conventional slicing makes predictions independently on each patch and then merges the detection boxes. LiG-DETR uses image patches as units for feature extraction that can be selected according to a budget. It zooms in on and encodes the selected patches, places their features back at their original image locations, and fills unselected regions with the corresponding whole-image features. This allows a single decoder to use both local details and whole-image context in one whole-image prediction pass.

Table 2 of the revised manuscript compares these two pipelines. With the same backbone and input resolution and comparable inference compute, LiG-DETR outperforms the RF-DETR-L slicing baseline by 1.6 percentage points in AP (37.1 vs. 35.5) and achieves approximately 2.1 times its end-to-end throughput (15.6 vs. 7.5 FPS).

R2. CPSR’s Distinct Mechanism and Experimental Evidence

Differences from prior methods. QueryDet uses low-resolution predictions to guide sparse detection on high-resolution features. Focus-DETR selects informative tokens in the encoder, and Dome-DETR’s MWAS uses a density map to select the feature windows that participate in attention. CPSR selects regions in image space according to a budget, zooms in on and re-encodes them to obtain additional local detail, and reassembles the resulting features at their original image locations to form a feature level aligned with the whole image.

Laplacian patch selection. Laplacian scoring is a replaceable patch selection strategy in CPSR. We evaluated 26,588 small objects on the VisDrone-DET2019 validation set, counting an object as covered if the center of its ground-truth box fell within a selected patch. When only 25% of the patches in each image are selected, Laplacian patch selection covers 48.3% of small objects, 1.93 times the expected coverage of uniform random selection under the same budget. When 50% and 75% of the patches are selected, coverage reaches 79.9% and 94.0%, exceeding the expected random coverage by 29.9 and 19.0 percentage points, respectively.

Laplacian score-based packing. In the full-image coverage setting with B=4B=4, we use the same model weights, the same 16 image patches, and the same encoding budget, varying only the packing layout at inference. Default score-based packing achieves 37.12 AP and 27.56 APₛ, differing from spatially adjacent packing’s 37.44 AP and 27.85 APₛ by only 0.32 and 0.29 percentage points, respectively. This shows that CPSR can pack patches from different parts of the original image into a single local input while maintaining detection accuracy comparable to that of spatially adjacent packing.

Global feature filling. The B=1B=1 configuration with G0G_0 filling in the main table achieves 34.6 AP and 25.3 APₛ, and the additional zero-filling test yields 22.9 AP and 16.5 APₛ, differences of 11.7 and 8.8 percentage points, respectively. By reusing whole-image features from the corresponding locations in G0G_0, CPSR can construct a reassembled feature level LL that covers the whole image while re-encoding only one-quarter of the image patches at high resolution.

R3. DA-AQA’s Budgeting Mechanism and Sensitivity

Differences from DQ-DETR and PAQI. DQ-DETR sets the number of queries based on the predicted object count range. PAQI first selects candidates by classification score, then uses a spatial density map to filter candidates and adjust the NMS threshold. The final query count is determined by the candidates that remain after filtering. DA-AQA first determines the query budget KqK_q from the candidate score distribution of each image, then uses the top-KqK_q candidate boxes to initialize the decoder’s query reference points. The entire budgeting process directly reuses the encoder’s existing candidate scores. In Table 6, the score-based packing configuration with DA-AQA achieves 37.1 AP, 1.6 percentage points higher than the three-level feature configuration without DA-AQA.

Rationale and scope of the cumulative score rule. For a given query count, selecting the highest-scoring candidates captures the largest possible sum of candidate scores. Therefore, when scores are accumulated in descending order, the number of candidates at which the cumulative score first reaches the fraction ρ\rho is the minimum raw budget needed to cover that fraction of the total score. The more concentrated the candidate scores, the fewer queries are needed to reach the same coverage. This rule measures the concentration of candidate responses, and its relationship to actual object density depends on candidate quality. For example, if the background candidates in an empty image all have similarly low scores, approximately ρM\rho M candidates are still needed to reach the threshold. The query bounds limit the maximum cost in such cases and ensure a minimum query capacity.

Hyperparameters and sensitivity. On VisDrone-DET2019, we use ρ=0.90\rho=0.90, a rounding step of r=32r=32, and query bounds of [300,500][300,500]. On AI-TOD-v2, we use ρ=0.70\rho=0.70, r=32r=32, and query bounds of [300,1500][300,1500]. We have listed these settings in the implementation details of the revised manuscript. For each dataset, the table below holds the model weights, rounding step, and query bounds fixed and varies only ρ\rho at inference.

Datasetρ\rhoMean query countAP (%)
VisDrone-DET2019 val0.60304.036.79
VisDrone-DET2019 val0.90500.037.12
AI-TOD-v2 test0.60678.031.52
AI-TOD-v2 test0.70856.531.70
AI-TOD-v2 test0.801,055.331.77

On VisDrone, lowering ρ\rho from 0.90 to 0.60 reduces the mean query count by 39.2%, with an AP drop of only about 0.3 percentage points. On AI-TOD-v2, ρ=0.70\rho=0.70 uses 18.8% fewer queries than ρ=0.80\rho=0.80, with AP only 0.07 percentage points lower. These results show that, in the tested settings, ρ\rho can substantially adjust the query budget while maintaining comparable detection accuracy.

R4. Comparisons at the Same Input Resolution and Comparable Inference Compute

Same input resolution. Most comparison methods use their default or commonly used 640×640640\times640 inputs, while RF-DETR-L and LiG-DETR both use the native 704×704704\times704 input resolution of RF-DETR-L. Table 2 of the revised manuscript adds the Dome-DETR-L results at 704×704704\times704. Dome-DETR-L achieves 36.0 AP at both resolutions. At the same 704×704704\times704 input resolution, LiG-DETR achieves 37.1 AP, a gain of 1.1 percentage points, and leads by 3.4 percentage points in AP for medium objects (49.9 vs. 46.5).

Comparable inference compute. At B=4B=4, LiG-DETR requires 645.7 GFLOPs and has 38.7M parameters, while whole-image RF-DETR-L requires 125.6 GFLOPs and has 33.9M parameters. To compare the two local processing pipelines under comparable compute budgets, we added an RF-DETR-L slicing baseline. For each image, both methods use the same backbone to encode one whole-image input and four local inputs, each at 704×704704\times704. At comparable computational cost (645.7 vs. 627.8 GFLOPs), LiG-DETR outperforms independent slicing by 1.6 AP (37.1 vs. 35.5) and 1.4 APₛ (27.6 vs. 26.2).

R5. End-to-End Latency, GPU Memory, and Efficiency Claims

We measured inference costs on the VisDrone-DET2019 validation set using the same RTX 3090 GPU.

MethodPreprocessing (ms)Model forward pass (ms)Postprocessing (ms)End-to-end latency (ms)Peak allocated GPU memory (MiB)
RF-DETR-L whole-image inference12.715.31.429.7200.9
RF-DETR-L slicing57.575.03.5132.6201.2
LiG-DETR (B=4B=4)14.248.41.764.0372.8

Per-stage timings and end-to-end latency were measured independently. Compared with the slicing baseline, LiG-DETR reduces end-to-end latency from 132.6 ms to 64.0 ms and increases throughput from 7.5 FPS to 15.6 FPS, approximately 2.1 times the baseline throughput. The measured time savings come mainly from preprocessing and the model forward pass, while peak GPU memory usage increases from 201.2 MiB to 372.8 MiB. We have clarified the scope of the comparisons supporting our efficiency claims in the abstract and main text.

R6. Additional Benchmarks and Reproducibility of DGDet-VisDrone

Table 5 of the revised manuscript adds horizontal bounding box detection results on the DIOR test set and the DOTA-v1.5 validation set. Compared with RF-DETR-L using the same backbone, LiG-DETR improves AP on DIOR from 58.6 to 59.3 and small-object AP from 18.7 to 21.1. On DOTA-v1.5, AP increases from 24.1 to 30.3, and medium-object AP increases from 22.3 to 32.7.

DGDet-VisDrone is built from annotated VisDrone-DET2019 images and contains 8,629 images. The Cloudy source domain is split at an approximately 2:1 ratio into 2,030 training images and 1,016 test images. Sunny, Dusk, Foggy, Night, and Rainy are used exclusively for evaluation on unseen domains. Table 1 of the manuscript reports the image count for each domain. LiG-DETR outperforms RF-DETR-L in AP on all five unseen domains, with gains ranging from 2.5 to 5.9 percentage points. We will release the code and processed data split files to facilitate reproduction of this evaluation protocol using the original VisDrone data.

R7. Method Presentation, Table Consistency, and Related Work

Method workflow and notation. We will improve the presentation of the method, clarify how the full-image coverage example relates to budget-based patch selection, and explain the notation consistently across the text, equations, and figures.

Dome-DETR-L results in the table and text. Thank you for pointing out the misaligned metrics in Table 2. We recalculated the metrics from the saved predictions and made the table and text consistent. At 640×640640\times640, Dome-DETR-L achieves 36.0 AP, 27.7 APₛ, 46.6 APₘ, and 62.6 APₗ. Table 2 of the revised manuscript also reports the new 704×704704\times704 comparison separately.

Relationship to non-uniform rescaling and sub-image feature stitching. We will discuss these two lines of work in the related work section and clarify how they differ from LiG-DETR. ZoomDet rescales the input image non-uniformly and maps the predicted boxes back to the original image after detection. LiG-DETR reassembles the local features obtained from zoomed-in encoding at their original image locations before detection. OV-Stitcher stitches sub-image features covering the whole image for semantic segmentation. LiG-DETR encodes only the patches selected within the budget at high resolution, fills unselected regions with whole-image features, and uses a single decoder to combine local details and whole-image context for detection.

R8. Per-Image Adaptation and Budget Selection for Deployment

The budget BB can be determined adaptively for each image. One possible extension is to generate proposals from whole-image features, aggregate the proposal scores by image patch, and then use these scores to determine BB and select patches for zoomed-in encoding before local encoding begins. The same decision rule can be used during training and inference, allowing the model to learn from reassembled features generated under different local encoding budgets during training.

We evaluate configurations with a fixed BB, so the number of local encoding passes per image can be determined before deployment. In practice, one can measure the end-to-end latency of each configuration on the target device and select the configuration and corresponding weights that achieve the highest detection accuracy among those that meet the latency requirement. For example, on VisDrone, B=3B=3 achieves 36.6 AP, only 0.5 percentage points below B=4B=4, while reducing inference compute from 645.7 to 522.9 GFLOPs, a reduction of approximately 19%.

Thanks for the authors' responses, which address some of my issues. I will keep my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.