Home

Awesome

Gliding vertex on the horizontal bounding box for multi-oriented object detection

This is the repository of paper "Gliding vertex on the horizontal bounding box for multi-oriented object detection".

image-20191218130556646

Install

The project is based on Maskrcnn Benchmark. Please follow its instructions to prepare the environment. For caution's sake, specific version 73ed879 is suggested to use. We list the some details here:

Python: 3.6.7
PyTorch: 1.2.0.dev20190704
CUDA: 9.0
CUDNN: 7

Compile the poly_nms:

REPO_ROOT/maskrcnn_benchmark/utils/poly_nms$ python setup.py build_ext --inplace

Don't forget to add maskrcnn_benchmark into $PYTHONPATH:

REPO_ROOT/maskrcnn_benchmark$ export PYTHONPATH=$PYTHONPATH:`pwd`

Run

Please edit the file maskrcnn_benchmark/config/paths_catalog.py to set the datasets.

Train:

REPO_ROOT$ python -m torch.distributed.launch --nproc_per_node=$NUM_GPUS tools/train_net.py --config-file $PATH_TO_CONFIG

Test:

REPO_ROOT$ python -m torch.distributed.launch --nproc_per_node=$NUM_GPUS tools/test_net.py --config-file $PATH_TO_CONFIG --ckpt=$PATH_TO_CKPT

If you want to train with your own data

This project use the json annotation file with COCO format. Make your directory layout like this:

.
└── trainset
    ├── images
    │   ├── 1.png
    │   └── 2.png
    └── labelTxt
        ├── 1.txt
        └── 2.txt

A example of the *.txt files ('1' means the object is difficult):

x1 y1 x2 y2 x3 y3 x4 y4 plane 0
x1 y1 x2 y2 x3 y3 x4 y4 harbor 1

Run the following Python snippet, and it will generate the json annotation file:

from txt2json import collect_unaug_dataset, convert
img_dic = collect_unaug_dataset( os.path.join( "trainset", "labelTxt" ) )
convert( img_dic, "trainset",  os.path.join( "trainset", "train.json" ) )

If you want to reproduce the results on DOTA

Config: configs/glide/dota.yaml

1. Prepare the data

Clone DOTA_Devkit as a sub-module:

REPO_ROOT$ git submodule update --init --recursive
REPO_ROOT/maskrcnn_benchmark/DOTA_devkit$ sudo apt-get install swig
REPO_ROOT/maskrcnn_benchmark/DOTA_devkit$ swig -c++ -python polyiou.i
REPO_ROOT/maskrcnn_benchmark/DOTA_devkit$ python setup.py build_ext --inplace

Edit the config.json and run:

REPO_ROOT$ python prepare.py

2. Train

You can train the model by yourself or directly use the snapshot provided by us (our method with FPN on DOTA): from Google Drive or Baidu NetDisk.

REPO_ROOT$ python -m torch.distributed.launch --nproc_per_node=3 tools/train_net.py --config-file configs/glide/dota.yaml

3. Test

REPO_ROOT$ python -m torch.distributed.launch --nproc_per_node=3 tools/test_net.py --config-file configs/glide/dota.yaml
# Edit ResultMerge.py and run it.
# srcpath = "REPO_ROOT/exp_dota/dota/inference/dota_test_cut/results"
REPO_ROOT/maskrcnn_benchmark/DOTA_devkit$ python ResultMerge.py

(optional) 4. Check if you use this project correctly

We provide the results on the valset in val_result.zip. If your produced results are different from them, maybe there is something wrong with your usage. The APs are in the following table. Note that in fact the results are meaningless since the valset has been used in training stage. It is just used for checking if you are using this project correctly.

PlaneBDBridgeGTFSVLVShipTCBCSTSBFRAHarborSPHCmAP
90.7589.1368.9479.8175.5484.5888.6790.9197.2590.5181.8288.3287.6183.1890.9185.86

BibTex:

@article{xu2019gliding,
  title={Gliding vertex on the horizontal bounding box for multi-oriented object detection},
  author={Xu, Yongchao and Fu, Mingtao and Wang, Qimeng and Wang, Yukang and Chen, Kai and Xia, Gui-Song and Bai, Xiang},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2020}
}

Please contact us if you have any question.