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

1 INTRODUCTION

• Aerial imagery spans large fields of view and object-scale and density variation, while small objects have weak cues vulnerable to clutter and feature compression.Effective detection therefore requires fine-grained local evidence alongside global context.Source: §1 · p. 1
• Image slicing improves local visibility but requires independent crop-level detection and box fusion, while feature- or query-level refinement works on already compressed full-image representations.Once fine-grained evidence is weakened by downsampling and compression, later enhancement may not fully recover it.Source: §1 · pp. 1–2
• Full-image encoding can leave small-object responses sparse or coarse, and higher-resolution features derived from that representation may not recover information already weakened.Locally magnified and reassembled features preserve clearer evidence before degradation.Source: §1 · p. 2, Figure 1 · p. 2
• LiG-DETR reassembles locally magnified features in global coordinates and jointly decodes them with global features using one DETR decoder.This treats image slices as feature-acquisition units rather than independent prediction units.Source: §1 · p. 2
• CPSR focuses high-resolution encoding on informative regions while retaining a dense feature layout, and DA-AQA adapts the query budget using encoder proposal scores.The two components reduce redundant computation along spatial and query dimensions, respectively.Source: §1 · p. 2

Image-Level Slicing for Aerial Small Objects

• Image-slicing methods improve local visibility but couple magnified crops to independent detection and subsequent box fusion or NMS-like post-processing.LiG-DETR uses crops for feature acquisition, then reassembles features in global coordinates for unified decoding.Source: §2 · p. 3

Multi-scale Feature and Query Modeling for Aerial Detection

• Multi-scale and DETR-based methods improve feature interaction, high-resolution processing, or query modeling, but generally enhance full-image representations.The related-work discussion notes that later refinement can hardly recover fine-grained information lost during early downsampling and compression.Source: §2 · p. 3
• LiG-DETR directly acquires high-fidelity local evidence before degradation and reintegrates it with global representations for unified end-to-end detection.This distinguishes its approach from feature- or query-level refinement of compressed full-image representations.Source: §2 · p. 3

3.1 OVERVIEW

• LiG-DETR jointly decodes global and reassembled local features with one DETR decoder, keeping prediction in the original image coordinate space.The shared encoder produces global features and locally magnified observations that are projected and reassembled according to their spatial locations.Source: §3.1 · p. 3, §3.2 · p. 3

3.2 GLOBAL-LOCAL REASSEMBLY FOR HIGH-FIDELITY LOCAL EVIDENCE

• Locally magnified crops provide a high-resolution feature level aligned with the full-image features for unified decoding.The full reassembly scheme divides the image into 2 × 2 crops, resizes them to encoder input size, and reassembles projected features at the resolution of G0.Source: §3.2 · p. 4

3.3 CONTEXT-PRESERVED SELECTIVE REASSEMBLY (CPSR)

• CPSR selects informative image regions using Laplacian-based scores, then fills unselected locations with G0 to preserve a dense, globally aligned feature layout.Selection operates directly on the image, retaining access to fine-grained cues before feature downsampling and compression.Source: §3.3 · p. 5

3.4 DENSITY-AWARE ADAPTIVE QUERY ALLOCATION (DA-AQA)

• DA-AQA sets an image-specific query budget from cumulative encoder proposal scores, using top-scoring proposal boxes to initialize decoder reference points.The budget reaches a predefined fraction of retained proposal-score mass, with rounding and clipping for stable, hardware-efficient computation.Source: §3.4 · pp. 5–6

3.5 TRAINING OBJECTIVE AND IMPLEMENTATION DETAILS

• LiG-DETR retains RF-DETR’s detection losses and original-image Hungarian matching, with no additional loss terms or local-branch supervision.Implementation uses RF-DETR-L with a pretrained, trainable DINOv2 encoder and evaluates four CPSR observation budgets.Source: §3.5 · p. 6

4.2 MAIN RESULTS

• 37.1% AP: LiG-DETRB=4 outperforms RF-DETR-L by 7.7 points on VisDrone-DET2019 and reaches the highest AP among compared methods.It improves APs from 18.6% to 27.6% and APm from 43.4% to 49.9%, while retaining 62.9% APl.Source: Table 2 · p. 7, §4.2 · pp. 7–8
• 31.8% overall AP: LiG-DETR leads the compared methods on AI-TOD-v2, outperforming RF-DETR-L by 6.9 points and improving APs and APm.APs rises from 23.6% to 30.8%, and APm from 45.4% to 49.8%, versus RF-DETR-L.Source: Table 3 · p. 7, §4.2 · p. 8

4.1 SETUP

• The evaluation covers VisDrone-DET2019, AI-TOD-v2, and DGDet-VisDrone; it reports COCO-style AP, object-size AP where specified, and efficiency measures.DGDet-VisDrone uses Cloudy as the source domain and five remaining domains as unseen targets; the comparison includes parameters, GFLOPs, and single-GPU FPS.Source: §4.1 · pp. 6–7

4.2 MAIN RESULTS

• 29.1% all-domain average AP: LiG-DETR exceeds RF-DETR-L by 5.0 points and improves across all five unseen DGDet-VisDrone target domains.Its all-domain average APs rises from 11.3% to 16.5% and APm from 33.3% to 39.2%, while Dome-DETR-L has higher APs.Source: Table 4 · p. 8, §4.2 · p. 8

4.3 ABLATION STUDY

• Adding reassembled local features to the global G1 baseline raises AP from 29.4% to 32.9%, but lowers APl from 65.1% to 59.0%.APs increases from 18.6% to 23.3%, indicating gains for small objects without the best balance across scales.Source: §4.3 · pp. 8–9, Table 5 · p. 8
• Without explicit fusion, keeping reassembled L as a separate feature level yields 35.5% AP, 47.8% APm, and 62.9% APl.Among global-feature combinations without local reassembly, {G0, G1} performs best at 30.4% AP and 20.2% APs; adding G2 gives no further gains.Source: §4.3 · p. 9
• DA-AQA raises AP from 35.5% to 37.1%; spatial-order and score-based CPSR packing perform nearly identically at 37.0% and 37.1% AP.Packing order has little effect once patch features are restored to their original spatial locations; score-based packing is the default.Source: §4.3 · p. 9
• With B=4, 16 patches achieve the best overall accuracy at 37.1% AP; finer partitioning to 64 patches lowers APs and APm.The reported GFLOPs are identical across granularities, while finer partitioning adds processing and reassembly overhead that lowers practical FPS.Source: §4.3 · p. 9, Table 7 · p. 9

5 CONCLUSION

• LiG-DETR reassembles high-resolution local evidence in global coordinates for unified end-to-end decoding, with CPSR selecting informative regions and query allocation adapting the decoder budget.The reported outcomes include improvements on small and medium objects, strong large-object performance, a favorable accuracy-efficiency trade-off, and improved cross-domain generalization.Source: §5 · p. 10

AI USE STATEMENT

• Generative AI assisted with translation, language editing, readability, and organization, but not research design, method development, data processing, or results interpretation.The authors state that they reviewed and revised the AI-assisted text and take responsibility for the final content.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

• The work uses publicly available aerial detection datasets or derived evaluation splits, and reports no new human-subject data or personally identifiable information.DGDet-VisDrone is derived from annotated VisDrone-DET2019 images.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

• The paper documents its architecture, training, evaluation, ablations, and DGDet-VisDrone domain splits, and says code and processed splits will be released.Source: REPRODUCIBILITY STATEMENT · p. 10

• On VisDrone-DET2019, LiG-DETR recovers more small and distant instances, especially in crowded traffic and nighttime scenes, with fewer obvious misses.Its predictions more closely follow ground-truth spatial distributions across varied scenes.Source: A QUALITATIVE ANALYSIS · p. 13, Figure 3 · p. 13
• The qualitative observations on both benchmarks are consistent with reported APs and APm improvements and support local high-resolution evidence as a complement to full-image representations.Source: A QUALITATIVE ANALYSIS · p. 13
• On AI-TOD-v2, LiG-DETR recovers weak isolated instances and densely distributed tiny objects, with predictions better matching the ground-truth distribution.The qualitative advantage is visible across different scene layouts.Source: A QUALITATIVE ANALYSIS · p. 13, Figure 4 · p. 14

• Increasing local-observation budget B improves accuracy but raises computation; B = 4 remains noticeably more expensive than the RF-DETR-L baseline.Further improving the efficiency of local high-resolution evidence acquisition is identified as future work.Source: B LIMITATIONS AND FUTURE WORK · p. 14
• CPSR uses a lightweight Laplacian-based high-frequency score, but those low-level cues are not always aligned with semantic object importance.The authors identify more effective allocation of high-resolution computation to informative regions as a future direction.Source: B LIMITATIONS AND FUTURE WORK · p. 14