FCOS: Fully Convolutional One-Stage Object Detection

The full paper is available at: https://arxiv.org/abs/1904.01355.

It performs well on DOTA data set

Required hardware

We use 4 Nvidia p100 GPUs. \

Installation

Testing-only installation

For users who only want to use FCOS as an object detector in their projects, they can install it by pip. To do so, run:

pip install torch  # install pytorch if you do not have it
pip install git+https://github.com/tianzhi0549/FCOS.git
# run this command line for a demo 
fcos https://github.com/tianzhi0549/FCOS/raw/master/demo/images/COCO_val2014_000000000885.jpg

Please check out here for the interface usage.

For a complete installation

This FCOS implementation is based on maskrcnn-benchmark. Therefore the installation is the same as original maskrcnn-benchmark.

Please check INSTALL.md for installation instructions. You may also want to see the original README.md of maskrcnn-benchmark.

Models

For your convenience, we provide the following trained models (more models are coming soon).

ResNe(x)ts:

All ResNe(x)t based models are trained with 16 images in a mini-batch and frozen batch normalization (i.e., consistent with models in maskrcnn_benchmark).

Model	AP (minival)
FCOS_imprv_R_101_FPN_2x	73.93
#Other tests to be tested

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/fcos_core/DOTA_devkit$ sudo apt-get install swig
REPO_ROOT/fcos_core/DOTA_devkit$ swig -c++ -python polyiou.i
REPO_ROOT/fcos_core/DOTA_devkit$ python setup.py build_ext --inplace

Edit the config.json and run:

REPO_ROOT$ python prepare.py

2 Training

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

3 Test

REPO_ROOT$ python tools/test_net.py --config-file configs/glide/dota.yaml --ckpt model

Contributing to the project

Any pull requests or issues are welcome.

Citations

Please consider citing our paper in your publications if the project helps your research. BibTeX reference is as follows.

@inproceedings{tian2019fcos,
  title   =  {{FCOS}: Fully Convolutional One-Stage Object Detection},
  author  =  {Tian, Zhi and Shen, Chunhua and Chen, Hao and He, Tong},
  booktitle =  {Proc. Int. Conf. Computer Vision (ICCV)},
  year    =  {2019}
}

lijain/FCOS-change