Source-linked AI summary
CLSC DETR: Reliable Candidate Ranking via Cross Layer Geometric Support for UAV Small Object Detection
Junyan Lin
TL;DR
UAV small-object detection suffers from unreliable candidate ranking because individual DETR queries may lack sufficient geometric evidence. CLSC-DETR aggregates cross-layer geometric support and calibrates classification with localization quality and classification reliability, improving AP and AP75 over MI-DETR on VisDrone and UAVDT.
Problem
Small, dense, and occluded UAV objects provide weak geometric evidence, making individual-query localization-quality estimation and candidate ranking unreliable.
Method
CLSC-DETR links final-layer queries with intermediate-layer candidates for geometric support and adaptively calibrates classification scores using localization quality and classification reliability.
Results
Compared with MI-DETR, CLSC-DETR improves AP/AP75 by 1.5%/2.0% on VisDrone and 3.0%/3.8% on UAVDT.
Takeaways & Limitations
Cross-layer geometric support and consistency calibration improve localization-quality estimation and candidate ranking for UAV small-object detection.
Abstract
from arXiv · showhide
Unmanned aerial vehicle (UAV) object detection is critical for applications such as target search, where accurate detection of small objects in complex aerial scenes remains challenging. The limited spatial extent, dense distribution, and frequent occlusion of small objects make reliable candidate ranking particularly difficult. Existing Detection Transformer (DETR) based methods improve ranking by estimating localization quality from individual queries and incorporating it into classification scores. However, a single query often lacks sufficient geometric evidence for small objects with weak boundary cues, resulting in unreliable quality estimation and unstable ranking. To address this limitation, we propose Cross Layer Local Support and Consistency Calibration for DETR, termed CLSC DETR. Specifically, the Cross Layer Local Support module establishes correspondences between final layer queries and intermediate layer candidates to aggregate complementary geometric evidence for more reliable localization quality estimation, while the Classification and Localization Consistency Calibration module adaptively adjusts classification scores according to localization quality and classification reliability to improve candidate ranking. Experiments show that CLSC DETR improves AP and AP$_{75}$ over the baseline by 1.5\% and 2.0\% on VisDrone, respectively, while achieving consistent improvements on UAVDT.
1. INTRODUCTION
UAV small-object detection is difficult because objects are small, dense, and occluded, leaving individual queries with insufficient geometric evidence for reliable localization-quality estimation and candidate ranking. CLSC-DETR addresses this with cross-layer geometric support and classification–localization consistency calibration, improving benchmark performance over MI-DETR.
- UAV imagery contains small, densely distributed, and heavily occluded objects, making reliable small-object detection essential for extracting target information.
- Existing DETR methods improve ranking by estimating localization quality, but individual candidates may lack sufficient geometric evidence for small objects.
- CLSC-DETR establishes correspondences between final-layer queries and intermediate-layer candidates to aggregate complementary geometric evidence for localization-quality estimation.
- The CLC module adaptively balances classification confidence and localization quality according to classification reliability, enabling more reliable candidate ranking.
- 1.5% AP and 2.0% AP75 improvements over the baseline are reported on VisDrone, with corresponding gains of 3.0% and 3.8% on UAVDT.
2. PROPOSED METHOD
CLSC-DETR improves candidate ranking by combining geometric evidence across encoder layers with adaptive classification–localization calibration. Its modules stabilize localization-quality estimation and align geometric quality with classification reliability.
- Overview: CLSC-DETR extends MI-DETR with Cross-Layer Local Support Quality Estimation and Classification–Localization Consistency Calibration.The two modules target small-object quality estimation and candidate ranking.
- Cross-Layer Local Support Quality Estimation: The CLS module searches local neighborhoods around final-layer candidates in intermediate layers and aggregates corresponding candidates using a shared prediction head.This provides cross-layer geometric evidence for final-layer queries.
- Cross-Layer Local Support Quality Estimation: Only the Top-K auxiliary candidates with the strongest geometric evidence are aggregated to suppress noisy correspondences.Strong support comes from candidates that highly overlap with the final-layer candidate.
- Cross-Layer Local Support Quality Estimation: The refined geometric quality is high only when intrinsic localization is reliable and cross-layer support is consistent.This reduces fluctuations caused by single-query prediction errors.
- Classification–Localization Consistency Calibration: The CLC module adaptively adjusts geometric-quality contributions using classification reliability estimated from maximum score, margin, and entropy-based certainty.A protection factor reduces geometric penalties when classification is more reliable than geometric quality.
3. EXPERIMENT AND ANALYSIS
Experiments on VisDrone and UAVDT show that CLSC-DETR improves detection performance, with complementary modules and adaptive calibration supporting candidate ranking and localization quality estimation.
- CLSC-DETR achieves the best overall performance on both VisDrone and UAVDT.
- 1.5%/1.5%/2.0%/1.1% gains in AP/AP50/AP75/APS over MI-DETR are reported on VisDrone.
- 3.0%/3.8%/3.8%/3.9% gains in AP/AP50/AP75/APS over MI-DETR are reported on UAVDT.
- Component Effectiveness: Combining CLS-Module and CLC-Module increases AP and AP75 by 1.5% and 2.0%, respectively.CLS-Module stabilizes localization quality estimation, while CLC-Module improves classification–localization consistency.
- Calibration Strategies: Adaptive calibration improves AP by 0.4% over both fixed-exponent alternatives.The strategy adjusts geometric-quality contributions according to classification reliability.
- Sensitivity Analysis: The best tested hyperparameter setting uses K = 2, α = 1.0, β = 0.20, and γ = 1.2.Excessive candidate aggregation or overly strong calibration may introduce noise and disturb classification–localization balance.
4. CONCLUSION
CLSC-DETR combines cross-layer geometric support with classification-reliability-aware calibration for UAV small-object detection. Compared with MI-DETR, it improves AP/AP75 by 1.5%/2.0% on VisDrone and 3.0%/3.8% on UAVDT.
- CLSC-DETR combines CLS and CLC modules to address localization-quality estimation bias and candidate-ranking instability.
- CLS uses cross-layer geometric support, while CLC calibrates scores according to classification reliability.
- The fixed local search and Top-K aggregation may not fully adapt to object-scale variations.Future work targets learnable neighborhoods and adaptive cross-layer aggregation.