fcos 0.9.3

Fully convolutional one-stage object detection (FCOS)

~Please note this is only a beta release at this stage~

FCOS: fully convolutional one-stage object detection

Fully convolutional one-stage object detection (FCOS) is a framework for per-pixel prediction of objects in images. FCOS doesn't rely on expensive anchor box calculations and their hyper-parameters, which is in contrast to state-of-the-art object detectors like RetinaNet, YOLOv3, and Faster R-CNN.

TODO: image of the system's output

This repository contains an open-source implementation of FCOS in Python, with access to pre-trained weights for a number of different models. The package provides PyTorch implementations for using training, evaluation, and prediction in your own systems. The package is easily installable with conda, and can also be installed via pip if you'd prefer to manually manage dependencies.

Our code is free to use, and licensed under BSD-3. We simply ask that you cite our work if you use FCOS in your own research.

Related resources

This repository brings the work from a number of sources together. Please see the links below for further details:

• our original paper: "FCOS: Fully convolutional one-stage object detection"
• our latest paper: "FCOS: A Simple and Strong Anchor-free Object Detector"
• the original FCOS implementation: https://github.com/tianzhi0549/FCOS
• implementation in the AdelaiDet toolbox: https://github.com/aim-uofa/AdelaiDet/blob/master/configs/FCOS-Detection/README.md#fcos-real-time-models

Installing FCOS

We offer three methods for installing FCOS:

1. Through our Conda package: single command installs everything including system dependencies (recommended)
2. Through our pip package: single command installs FCOS and Python dependences, you take care of system dependencies
3. Directly from source: allows easy editing and extension of our code, but you take care of building and all dependencies

Conda

The only requirement is that you have Conda installed on your system, and NVIDIA drivers installed if you want CUDA acceleration. We provide Conda packages through Conda Forge, which recommends adding their channel globally with strict priority:

conda config --add channels conda-forge
conda config --set channel_priority strict

Once you have access to the conda-forge channel, FCOS is installed by running the following from inside a Conda environment:

u@pc:~$ conda install fcos

We don't explicitly lock the PyTorch installation to a CUDA-enabled version to maximise compatibility with our users' possible setups. If you wish to ensure a CUDA-enabled PyTorch is installed, please use the following installation line instead:

u@pc:~$ conda install pytorch=*=*cuda* fcos

You can see a list of our Conda dependencies in the FCOS feedstock's recipe.

Pip

Before installing via pip, you must have the following system dependencies installed if you want CUDA acceleration:

• NVIDIA drivers
• CUDA

Then FCOS, its custom CUDA code, and all of its Python dependencies, can be installed via:

u@pc:~$ pip install fcos

From source

Installing from source is very similar to the pip method above, accept we install from a local copy. Simply clone the repository, enter the directory, and install via pip:

u@pc:~$ pip install -e .

Note: the editable mode flag (-e) is optional, but allows you to immediately use any changes you make to the code in your local Python ecosystem.

We also include scripts in the ./scripts directory to support running FCOS without any pip installation, but this workflow means you need to handle all system and Python dependencies manually.

Using FCOS

TODO

FCOS from the command line

FCOS Python API

Citing our work

If using FCOS in your work, please cite our original ICVV paper:

@inproceedings{tian2019fcos,
  title={FCOS: Fully convolutional one-stage object detection},
  author={Tian, Zhi and Shen, Chunhua and Chen, Hao and He, Tong},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={9627--9636},
  year={2019}
}

Or our more recent TPAMI journal with further details of our work:

@article{tian2021fcos,
  author={Tian, Zhi and Shen, Chunhua and Chen, Hao and He, Tong},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  title={FCOS: A Simple and Strong Anchor-free Object Detector},
  year={2020},
  volume={},
  number={},
  pages={1-1},
  doi={10.1109/TPAMI.2020.3032166}}