N^Segment: Label-specific Deformations for Remote Sensing Image Segmentation

Yechan Kim*, DongHo Yoon*, SooYeon Kim, and Moongu Jeon

This repo includes:

• Official implementation of our proposed approach (NSegment)

Announcement:

• Aug. 2025: Our pilot work has been systematically extended and enhanced! Visit NSegment+!
• Aug. 2025: Our paper is accepted to IEEE Geoscience and Remote Sensing Letters! 🎉
• Apr. 2025: Our preprint is now available at https://arxiv.org/abs/2504.19634.
• Apr. 2025: We have released the official code of our proposed approach!

Overview:

• With our strategy, you can boost remote sensing image segmentation.

Our algorithm in Python3 (with OpenCV, Numpy, and MMSegmentation)

import cv2
import numpy as np

from mmseg.registry import TRANSFORMS


@TRANSFORMS.register_module()
class NoisySegment:
    def __init__(self, alpha_sigma_list=None, prob=0.5):
        if not alpha_sigma_list:
            alpha_sigma_list = [(alpha, sigma) for alpha in [1, 15, 30, 50, 100] 
                                 for sigma in [3, 5, 10]]
        self.alpha_sigma_list = alpha_sigma_list
        self.prob = prob

    def __call__(self, results):
        if np.random.rand() > self.prob:
            return results
        segment = results['gt_seg_map']
        noisy_segment = self.transform(segment)
        results['gt_seg_map'] = noisy_segment
        return results

    def transform(self, segment, random_state=None):
        if random_state is None:
            random_state = np.random.RandomState(None)
        alpha, sigma = self.alpha_sigma_list[random_state.randint(0, len(self.alpha_sigma_list))]
        shape = segment.shape[:2]
        dx = alpha * (2 * random_state.rand(*shape) - 1)
        dy = alpha * (2 * random_state.rand(*shape) - 1)
        dx = cv2.GaussianBlur(dx, (0, 0), sigma)
        dy = cv2.GaussianBlur(dy, (0, 0), sigma)
        x, y = np.meshgrid(np.arange(shape[1]), np.arange(shape[0]))
        map_x = np.clip(x + dx, 0, shape[1] - 1).astype(np.float32)
        map_y = np.clip(y + dy, 0, shape[0] - 1).astype(np.float32)
        noisy_segment = cv2.remap(segment, map_x, map_y, interpolation=cv2.INTER_NEAREST)
        return noisy_segment

How to use:

• If you use our git repository, our augmentation method is already included and registered.
• Simply add our NoisySegment to train_pipeline in your model configuration file. Below is an example:

  train_pipeline = [
      dict(type='LoadImageFromFile'),
      dict(type='LoadAnnotations'),
      dict(type='NoisySegment'), # ⭐
      ...
  ]

• For better training, you might adjust various $(\alpha, \sigma)$ parameter pairs via alpha_sigma_list as follows:

  dict(type='NoisySegment', alpha_sigma_list=[(1, 3), (1, 5) ...])

• If you want our method be applied with a prob of 80%? (Note: default prob = 0.5)

  dict(type='NoisySegment', alpha_sigma_list=[(1, 3), (1, 5) ...], prob=0.8)

Preliminaries:

• Step 1. Create a conda environment with Python 3.8 and activate it.

  conda create -n nsegment python=3.8 -y
  conda activate nsegment

• Step 2. Install PyTorch with TorchVision following official instructions. The below is an example.

  conda install pytorch==2.1.2 torchvision==0.16.2 pytorch-cuda=12.1 -c pytorch -c nvidia

• Step 3. Install MMSegmentation (v1.2.2) (v1.2.2 is the latest version of 2024).

  # ⚠️ No need to clone MMSeg (e.g. "git clone https://github.com/open-mmlab/mmsegmentation; rm -rf mmsegmentation/.git"). Already cloned! 
  pip install -U openmim==0.3.9
  mim install mmengine==0.10.7
  mim install mmcv==2.1.0
  pip install -v -e mmsegmentation/
  pip install ftfy==6.2.3
  pip install regex==2024.11.6

Citation:

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

• For Latex:

  @article{kim2025nsegment,
    title={NSegment: Label-specific Deformations for Remote Sensing Image Segmentation},
    author={Kim, Yechan and Yoon, DongHo 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, DongHo Yoon, SooYeon Kim, and Moongu Jeon. "NSegment: Label-specific Deformations for Remote Sensing Image Segmentation." IEEE Geoscience and Remote Sensing Letters 22 (2025).

We also highly recommend referring to our prior study as follows:

• GitHub: https://github.com/unique-chan/NBBOX
• 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.