Detection-of-Multiclass-Objects-in-Optical-Remote-Sensing-Images

Detection of Multiclass Objects in Optical Remote Sensing Images GRSL paper

Running Environments:

Ubuntu 16.04 Anaconda(Anaconda3-5.0.1-Linux-x86_64 python 3.6)
torch 0.4.0a Pytorch

  git clone -o 7b31d33e80afef71af8a61569c24ce4df485b621 --recursive https://github.com/pytorch/pytorch
  git checkout -b 040 7b31d33e80afef71af8a61569c24ce4df485b621
  or(git checkout -b 040 040336f5dc5aa8b80fd8fc6f96892dd823b22752)
  git submodule update --init
  python setup.py install

torchvision Vision

Prerequisites

  pip install shapely
  sudo apt-get install ttf-mscorefonts-installer

Download weight parameters form Google Drive and put it into backup file folder.
Download DOTA dataset

Note: Please config the improved orientated response network follows IORN

Test:

  python detect.py  

Train:

  All training images of task 2 are divided into 1024×1024 pixel patches by the DOTA development kit  

  we can delete the patches which contain no object.
  Generate train list of all divided images like "train_img_example/train.txt"   

  Change train list path in "cfg/dota.data"  

  python train.py cfg/dota.data cfg/orn_4_dota.cfg backup/000057.weights  

Evaluate for DOTA testset or valset:

  Generate test image list like "test_img/test_img_list.txt"  

  Run "python test_for_map.py"

  For Task2 valset, we can evaluate by DOTA development kit.
  For Task2 testset, we must registrate and submit on the Horizontal Evaluation Server.

NOTE:1, This project is built based on the code marvis. Thanks all the contributors. MIT License. 2, DOTA dataset was updated, we have not tested the new version. 3, If you can't find coresponding version of running environments, please contact me. 4, If you have any suggestions or questions, please contact me liuwenchaomuc@163.com.