Keykeykeykeykeykey/RT-DETR-IELC3-SHSA: test

Official PyTorch implementation of the paper:

Robust Object Detection in Optical Remote Sensing Imagery via Partial-Channel Self-Attention and Intensity-Enhanced Feature Fusion

Overview

Optical remote sensing object detection remains challenging due to sparse object distribution, large scale variation, and complex background interference. To address these issues, we propose AIRT-Det, a lightweight CNN–Transformer hybrid detector built upon the RT-DETR framework.

AIRT-Det introduces two complementary modules:

SHSAIFI

Single-Head Self-Attention-based Internal Feature Interaction

• Performs self-attention on only a subset of informative feature channels.
• Reduces redundant global interactions.
• Improves contextual representation efficiency.
• Maintains low computational overhead.

IFusion

Intensity Enhance Fusion

• Introduces intensity-regulated nonlinear feature modulation.
• Stabilizes cross-scale feature interaction.
• Enhances discriminative object responses.
• Improves multi-scale feature fusion for small and low-contrast targets.

The proposed framework achieves superior detection performance while preserving real-time inference capability.

Framework

<p align="center"> <img src="assets/framework.png" width="900"> </p>

The overall architecture consists of:

1. CNN Backbone Feature Extraction
2. SHSAIFI-enhanced Transformer Encoder
3. IFusion-based Multi-scale Feature Fusion
4. RT-DETR Decoder

Key Features

• End-to-End Object Detection
• Real-Time Inference
• Lightweight Design
• Efficient Global Context Modeling
• Adaptive Multi-scale Feature Fusion
• State-of-the-Art Performance on Remote Sensing Benchmarks

Experimental Results

DOTA-v1.5

| Method | mAP (%) | |----------|----------| | RT-DETR | 67.1 | | AIRT-Det | 72.5 |

HRSC2016

| Method | mAP (%) | |----------|----------| | RT-DETR | 91.0 | | AIRT-Det | 91.7 |

DIOR

| Method | mAP (%) | |----------|----------| | RT-DETR | XX.X | | AIRT-Det | XX.X |

Installation

Environment

• Python 3.11
• PyTorch 2.4.1
• CUDA 12.4
• cuDNN 7.5.0

Clone Repository

git clone https://github.com/yourname/AIRT-Det.git
cd AIRT-Det

Create Conda Environment

conda create -n airtdet python=3.11
conda activate airtdet

Install Dependencies

pip install -r requirements.txt

Dataset Preparation

DOTA-v1.5

Download:

https://captain-whu.github.io/DOTA/

Directory structure:

datasets/
└── DOTA-v1.5/
    ├── images/
    ├── labels/
    └── split/

HRSC2016

datasets/
└── HRSC2016/
    ├── images/
    ├── annotations/
    └── labels/

DIOR

datasets/
└── DIOR/
    ├── JPEGImages/
    ├── Annotations/
    └── labels/

Training

Train AIRT-Det on DOTA-v1.5:

python train.py \
    --model rtdetr-airtdet.yaml \
    --data dota.yaml \
    --epochs 300 \
    --batch 2 \
    --imgsz 640

Evaluation

python val.py \
    --weights runs/train/best.pt \
    --data dota.yaml

Inference

python predict.py \
    --weights best.pt \
    --source demo.jpg

Ablation Study

HRSC2016

| SHSAIFI | IFusion | mAP (%) | |----------|----------|----------| | ✗ | ✗ | 91.0 | | ✓ | ✗ | 91.2 | | ✗ | ✓ | 91.2 | | ✓ | ✓ | 91.7 |

DOTA-v1.5

| SHSAIFI | IFusion | mAP (%) | |----------|----------|----------| | ✗ | ✗ | 67.1 | | ✓ | ✗ | 68.1 | | ✗ | ✓ | 70.0 | | ✓ | ✓ | 72.5 |

Repository Structure

AIRT-Det/
├── configs/
│   └── rtdetr-airtdet.yaml
├── models/
│   ├── SHSAIFI.py
│   ├── IFusion.py
│   └── IELC3.py
├── datasets/
├── scripts/
├── weights/
├── assets/
│   └── framework.png
├── train.py
├── val.py
├── predict.py
├── requirements.txt
└── README.md

Code Availability

The source code associated with this study is publicly available at:

https://github.com/Keykeykeykeykeykey/RT-DETR-IELC3-SHSA

Archived version (DOI):

https://doi.org/xxxxxxxx

Citation

If you find this work useful, please cite:

Acknowledgements

This work is built upon the following excellent open-source projects:

• RT-DETR
• Ultralytics
• PyTorch

We sincerely thank the authors for making their code publicly available.

License

This project is released under the MIT License.

See the LICENSE file for details.