Light-Weight Semantic Segmentation Network for UAV Remote Sensing Images

Briefly

• This repo introduces a light-weight semantic segmentation network for UAV Remote Sensing Images

• The network only requires 9M parameters

• The experiments on the ISPRS Vaihingen dataset, UAVid dataset and UDD6 dataset had verify the effectiveness of it.

Environment

Runtime environment

• Ubuntu 16.04
• PyTorch 1.6.0
• CUDA10.1+
• Nvidia GTX2080Ti

Models

• All the models involved in models/

• Under the condition of the image size is 512x512, the performances of our models on the Vaihingen dataset are as follows:

  Model	mF1	mIoU	OA	Params(M)
  LWN	86.79	77.11	88.27	9
  LWN-A	87.62	78.38	88.85	15

  UAVid:

  Model	mIoU	OA	Params(M)
  LWN	67.82	87.13	9
  LWN-A	69.02	87.66	15

  UDD:

  Model	mF1	mIoU	OA	Params(M)
  LWN	86.19	76.78	88.75	9
  LWN-A	86.79	77.19	88.93	15

​

Training

• It is recommended to make a new dir named data and save or link the dataset under it.

• Images and labels are recommended to crop to 512*512

• Then prepare the data as follows:

• data/uavid
  |-- train
  |   |-- image
  |   |   |-- seq1_000000.png
  |   |   |-- ...
  |   |-- label
  |   |   |-- seq1_000000.png
  |   |   |-- ...
  |-- val
  |   |-- image
  |   |   |-- seq16_000000.png
  |   |   |-- ...
  |   |-- label
  |   |   |-- seq16_000000.png
  |   |   |-- ...
  

• Then set the parameters for training phase, such as dataset, model_type , data_root and learning rateon config.ini.

• python main.py