Introduction

Anchor optimization for RetinaNet based on Improving RetinaNet for CT Lesion Detection with Dense Masks from Weak RECIST Labels by Martin Zlocha, Qi Dou and Ben Glocker. Implementation makes heavy use of keras-retinanet.

For questions and discussion join the Keras Slack and either message me directly (username: martinzlocha) or join the #keras-retinanet channel.

Setup

1. Clone this repository
2. pip install .
3. python setup.py build_ext --inplace

Usage

Basic usage:

1. Define your own dataset in a csv format, for more information follow the guide in keras-retinanet.
2. Run anchor-optimization PATH_TO_CSV

Additional options:

• scales and ratios parameters allows you to define the number scales and ratios in your anchor configuration.
• objective parameter allows you to specify how the objective function is calculated. The avg parameter optimizes the average overlap between the anchors and the ground truth. We suggest that you use the focal parameter (default) instead because it ensures that there are less ground truth lesions with lower than 0.5 IoU with an anchor.
• popsize parameter allows you to specify the population multiplier used by differential evolution. Higher values result in better results however the computation will also be slower.
• By default the images and the detected objects are resized using the same method as in keras-retinanet implementation. If you do not wish to resize the image use the no-resize flag or specify different image-min-side and image-max-side.
• By default we ignore strides when optimizing the anchor configuration, this makes it feasible to optimize the anchors for a large number of objects in a short time. We have recently added an include-stride flag. This makes the computation much slower however it is more accurate. We suggest you only use it if your dataset is small.
• threads to allow parallelization.

To reproduce our results:

anchor-optimization PATH_TO_CSV --ratios=5 --no-resize

Notes

• This repository has been tested on python 3.7.
• This repository relies on keras-retinanet 0.5.1.

Contributions to this repository are welcome.

Citation

If you find the code or the optimized anchors useful for your research, please consider citing our paper.

@article{zlocha2019improving,
  title={Improving RetinaNet for CT Lesion Detection with Dense Masks from Weak RECIST Labels},
  author={Zlocha, Martin and Dou, Qi and Glocker, Ben},
  journal={arXiv preprint arXiv:1906.02283},
  year={2019}
}