D³R-DETR - Tiny Object Detection

D³R-DETR: DETR with Dual-Domain Density Refinement for Tiny Object Detection in Aerial Images
Z Wen, Z Yang, X Bao, L Zhang, X Xiang, W Li, Y Liu
arXiv preprint, 2026
arXiv

Project Overview

Challenge: Tiny Object Detection in Aerial Images

Aerial image object detection presents significant challenges:

• Extremely small object sizes (often < 16×16 pixels)
• High object density in urban scenes
• Large scale variations across the image
• Complex backgrounds with cluttered environments

Solution: Dual-Domain Density Refinement

D³R-DETR introduces:

1. Dual-domain feature refinement (spatial + frequency)
2. Density-aware query generation
3. Multi-scale deformable attention

Key Innovations

1. Dual-Domain Feature Refinement

Combines spatial and frequency domain information:

• Spatial domain: Local texture and shape features
• Frequency domain: Global structural information
• Adaptive fusion for optimal representation

2. Density-Aware Query Generation

Handles varying object densities:

• Density map estimation
• Adaptive query distribution
• Region-specific processing

3. Enhanced Deformable Attention

Improved attention mechanism:

• Multi-scale feature aggregation
• Deformable sampling for small objects
• Efficient computation

Method

Architecture Overview

Aerial Image
    ↓
[Backbone Feature Extraction]
├── Multi-scale features
└── Feature pyramid
    ↓
[Dual-Domain Refinement]
├── Spatial refinement
├── Frequency refinement
└── Domain fusion
    ↓
[Density-Aware Detection]
├── Query generation
├── Deformable attention
└── Box prediction
    ↓
Detection Results

Results

Performance on aerial image datasets:

• State-of-the-art tiny object detection
• Improved detection for small targets
• Robust to varying object densities

Applications

• Urban planning and monitoring
• Traffic flow analysis
• Disaster damage assessment
• Environmental monitoring

Citation

@article{wen2026d3rdetr,
  title={D³R-DETR: DETR with Dual-Domain Density Refinement for Tiny Object Detection in Aerial Images},
  author={Wen, Z and Yang, Z and Bao, X and Zhang, L and Xiang, X and Li, W and Liu, Y},
  journal={arXiv preprint arXiv:2501.02747},
  year={2026}
}