ChainerCV is a collection of tools to train and run neural networks for computer vision tasks using Chainer.

You can find the documentation here.

Supported tasks:

• Image Classification (ResNet, SENet, VGG)
• Object Detection (tutorial, Faster R-CNN, FPN, SSD, YOLO, Light-Head R-CNN)
• Semantic Segmentation (SegNet, PSPNet, DeepLab v3+)
• Instance Segmentation (FCIS, Mask R-CNN)

ChainerCV is developed under the following three guiding principles.

• Ease of Use -- Implementations of computer vision networks with a cohesive and simple interface.
• Reproducibility -- Training scripts that are perfect for being used as reference implementations.
• Compositionality -- Tools such as data loaders and evaluation scripts that have common API.

$ pip install -U numpy
$ pip install chainercv

The instruction on installation using Anaconda is here (recommended).

• Chainer and its dependencies
• Pillow
• Cython (Build requirements)

For additional features

• Matplotlib
• OpenCV
• SciPy
• mpi4py
• pycocotools

Environments under Python 2.7.12 and 3.6.0 are tested.

• The master branch is designed to work on Chainer v6 (the stable version) and v7 (the development version).
• The following branches are kept for the previous version of Chainer. Note that these branches are unmaintained.
  • 0.4.11 (for Chainer v1). It can be installed by pip install chainercv==0.4.11.
  • 0.7 (for Chainer v2). It can be installed by pip install chainercv==0.7.
  • 0.8 (for Chainer v3). It can be installed by pip install chainercv==0.8.
  • 0.10 (for Chainer v4). It can be installed by pip install chainercv==0.10.
  • 0.12 (for Chainer v5). It can be installed by pip install chainercv==0.12.
  • 0.13 (for Chainer v6). It can be installed by pip install chainercv==0.13.

• Image
  • The order of color channel is RGB.
  • Shape is CHW (i.e. (channel, height, width)).
  • The range of values is [0, 255].
  • Size is represented by row-column order (i.e. (height, width)).
• Bounding Boxes
  • Shape is (R, 4).
  • Coordinates are ordered as (y_min, x_min, y_max, x_max). The order is the opposite of OpenCV.
• Semantic Segmentation Image
  • Shape is (height, width).
  • The value is class id, which is in range [0, n_class - 1].

These are the outputs of the detection models supported by ChainerCV.

If ChainerCV helps your research, please cite the paper for ACM Multimedia Open Source Software Competition. Here is a BibTeX entry:

@inproceedings{ChainerCV2017,
    author = {Niitani, Yusuke and Ogawa, Toru and Saito, Shunta and Saito, Masaki},
    title = {ChainerCV: a Library for Deep Learning in Computer Vision},
    booktitle = {ACM Multimedia},
    year = {2017},
}

The preprint can be found in arXiv: https://arxiv.org/abs/1708.08169