ALOD_DML

This repo contains the codes for the TGRS paper: "Minimizing Sample Redundancy for Label-efficient Object Detection in Aerial Images"

Introduction

In this paper, we propose a novel labeling pattern that acquires heterogeneous object labels in a class-orthogonal manner, called Attribute-aware labeling pattern for Label-efficient Object Detection in aerial images (ALOD).

Meanwhile, we also propose a new learning pipeline under the ALOD labeling pattern, called Dynamic Multiview Learning (DML)

Requirements

• Python 3.7+
• PyTorch 1.13.1
• mmcv-full 1.7.0
• mmdet 2.25.3 (https://github.com/open-mmlab/mmdetection)
• mmrotate 0.3.3 (https://github.com/open-mmlab/mmrotate)
• CUDA 11.3+ (recommended for GPU acceleration)

Installation

cd thirdparty/mmdetection/
pip install -v -e .
cd ../mmrotate/
pip install -v -e .
cd ../../
pip install -v -e 

Dataset Preparation

1. Dataset Structure

First, follow the DOTA data format in MMrotate to crop large images. The final dataset folder structure should be:

data/
└── dota1/
    ├── train_obb/
    │   └── split_images/
    │       ├── annfiles/  # Annotation files (*.txt)
    │       └── images/    # Cropped image files
    └── val_obb/
        └── split_images/
            ├── annfiles/
            └── images/

2. Data Processing Scripts

Run the following commands to prepare the K-ALOD dataset and build the Reliable Samples Pool:

# Step 1: Construct K-ALOD dataset from raw DOTA data
python prepare_data_and_model/1_conduct_K-ALOD_dataset.py

# Step 2: Crop object patches to build Reliable Samples Pool
python prepare_data_and_model/2_cut_image_patches.py

Pretraining

Before formal training, perform pretraining with the base detector and convert the model to T-S mode:

# Step 1: Pretrain the base detector (using 1 GPU, adjust GPU number as needed)
cd thirdparty/mmrotate
./tools/dist_train.sh configs/rotated_faster_rcnn/rotated_faster_rcnn_r50_fpn_for_alod.py 1
# Step 2: Convert the pretrained base model to T-S (Teacher-Student) mode
 
cd ../../
python prepare_data_and_model/generated_the_pretrained_T-S_model.py

Training

Use the following command to start the formal training of the ALOD-DML model:

# Train with 2 GPUs (adjust GPU number according to your device)
bash ./tools/dist_train.sh configs/K_ALOD_dotav1/PLOD_v1_rofaster_dota1_1ins.py 2

Quantitative Results

The table above shows the quantitative performance of our ALOD-DML method compared with other state-of-the-art label-efficient object detection methods on the DOTA dataset.

Visualization

The figure above visualizes the detection results of our ALOD-DML method on aerial images, demonstrating its effectiveness in recognizing objects with different orientations and sizes.

Citation

If you use this code or our work in your research, please cite our TGRS paper:

@article{ALOD_DML,
  author={Zhang, Ruixiang and Xu, Chang and Zhu, Haoran and Xu, Fang and Yang, Wen and Zhang, Haijian and Xia, Gui-Song},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Minimizing Sample Redundancy for Label-Efficient Object Detection in Aerial Images}, 
  year={2025},
  volume={63},
  number={},
  pages={1-14},
  doi={10.1109/TGRS.2025.3562395}
}

Acknowledgements

This project is built upon the following open-source libraries, and we sincerely appreciate their contributions:

• MMdetection: OpenMMLab object detection toolbox
• MMrotate: OpenMMLab rotated object detection toolbox