The Astrophysical CIrcles Detector (ACID) is a computer vision package trained to detect quasi-circular objects in, non-exclusively, astrophysical images. Such objects include impact craters on any well imaged solar system object, eddies (cyclones) on gas planets, boulders on asteroids and comets, and even HI holes in galaxies.
- It is an ensemble model built on top of the MaskRCNN semantic segmentation framework (Matterport implementation).
- Easy to install and use. No prior experience with machine learning is needed. Can be used on a personal computer.
- Trained on a massive and highly augmented craters-only dataset. Detection of any other object is a form of transfer-learning.
- Returns the location, size, and shape of the detected objects.
- Packs many convenience functions to preprocess images and postprocess results.
- Still in active development.
ACID can be downloaded and used right away. No installation is needed. It does however have many dependencies that must be installed. A conda environment with all of the needed packages can be installed using the provided yml file : conda env create -f acid_env.yml
Finally the models weights need to be downloaded from https://nyu.box.com/s/n7g419f46ib1bc8lmdelafn6w928jx77 and extracted into ./models/ .
Multiple usage examples are included as jupyter notebooks and python scripts. Users are advised to start with example_moon as it is the most detailed.
A dedicated ACID paper will be published soon. In the meanwhile, if you find ACID useful, please cite this Github repository and the following two papers:
Ali-Dib, M. et al. (2020) Icarus, Volume 345, article id. 113749.
Silburt, A., Ali-Dib, M. et al. (2019) Icarus, Volume 317, p. 27-38.
The MIT License (MIT)