NBBOX: Noisy Bounding Box Improves Remote Sensing Object Detection

Yechan Kim, SooYeon Kim, and Moongu Jeon

This repo includes:

• Official implementation of our proposed approach

Announcement:

• Apr. 2025: Our pilot work has been extended for remote sensing image segmentation! Visit NSegment!
• Feb. 2025: Erratum: In our experiments, we used Swin-S instead of Swin-T; however, there was a typo in the manuscript where we mistakenly wrote Swin-T. We hope you understand. 🙏
• Jan. 2025: Our paper is accepted to IEEE Geoscience and Remote Sensing Letters! 🎉
• Dec. 2024: We have released the official code of our proposed approach!
• Sep. 2024: Our preprint is now available at https://arxiv.org/abs/2409.09424.

Overview:

• With our noisy bounding box transformation, you can boost remote sensing object detection.

Preliminaries:

• Install all necessary packages listed in the requirements.txt.
• Simply add our NoisyBBOX to train_pipeline in your model configuration file. Below is an example:

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(type='NoisyBBOX'), # Our transformation (⭐) should be placed directly after 'LoadAnnotations'
    ...
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]

Training via our strategy:

• Our NoisyBBOX has the following arguments to control:
  • scale_range (min, max) (dtype: float): Scale each bounding box by a random number between min and max, during training. Each box is affected by a different random number.
    • isotropically_rescaled (dtype: boolean): If this variable is True, each box is rescaled while preserving the aspect-ratio. Otherwise, not.
  • angle_range (min, max) (dtype: int): Rotate each bounding box by a random number between min and max, during training. Each box is affected by a different random number.
  • translate_range (min, max) (dtype: int): Translate each bounding box by a random number between min and max, during training. Each box is affected by a different random number.
    • isotropically_translated (dtype: boolean): If this variable is True, each box is translated while preserving the aspect-ratio. Otherwise, not.
  • threshold (dtype: float): If the width or height of a bounding box is below a certain threshold, transformation is not applied.
• When training with this Git repo, you can easily control the arguments as follows:
  • Example: { scale_range: (0.7, 1.0), isotropically_rescaled: True }, angle_range: (-2, 2), { translate_range: (-1, 1), isotropically_translated: True }

  python train.py --model-config "my_src/my_cfg/rotated_faster_rcnn_r50_fpn_1x_dior_le90.py" \
                  --data-config "my_src/my_cfg/for_dior/cfg_for_data_DIOR_obb.py" \
                  --train-config "my_src/my_cfg/for_dior/cfg_for_train_DIOR_obb.py" \
                  --dataset-class "my_src/my_dataset/dior_obb.py" --batch-size 8   \
                  --data-root "DIOR" --epochs 1 --work-dir "work_dirs/dior_faster_r50" --gpu-id 0 --obb --tag "noisy-bbox" --init_weights \
                  --scale_min 0.7 --scale_max 1.0 --isotropically_rescaled \
                  --angle_min -2 --angle_max 2 \
                  --translate_min -1 --translate_max 1 --isotropically_translated // ⭐
  

• If you hope to train detectors with your own MMDetection/Rotate-based codes, please follow steps below:
  • Copy and paste my_transform/noisy_bbox.py into your code. Ensure that our transform to be registered in MMDetection/Rotate's PIPELINES.
  • Also, add our transform with parameters to train_pipeline in your model configuration file as follows:

  train_pipeline = [
    ...
    dict(type='NoisyBBOX', scale_range=(0.7, 1.0), 
         isotropically_rescaled=True, angle_range=(-2, 2), ...),
    ...
  ]

Citation:

If you use this code for your research, please cite the following paper:

• For Latex:

  @article{kim2025nbbox,
    title={NBBOX: Noisy Bounding Box Improves Remote Sensing Object Detection},
    author={Kim, Yechan and Kim, SooYeon and Jeon, Moongu},
    journal={IEEE Geoscience and Remote Sensing Letters},
    volume={22},
    year={2025},
    publisher={IEEE}
  }

• For Word (MLA Style):

  Yechan Kim, SooYeon Kim, and Moongu Jeon. "NBBOX: Noisy Bounding Box Improves Remote Sensing Object Detection." IEEE Geoscience and Remote Sensing Letters 22 (2025).

Contribution:

If you find any bugs or have opinions for further improvements, please feel free to create a pull request or contact me (yechankim@gm.gist.ac.kr). All contributions are welcome.