The network (a normalizing flow) takes in density maps instead of pictures of anomalies. These are conditioned on features extracted from the corresponding aerial images. The density maps were produced from the centre points of a manually curated dataset of 6,000 images, with approximately 400 of these having object annotations.
When running this repository, there are 5 folders that will store outputs. VIZ stores images of density map reconstructions. RUNS stores tensorboard logging. MODELS and WEIGHTS saves the model objects and state dictionaries for trained models. CSTATE stores copies of the config.py file used to define the parameters of the hyper parameters of each model and general training options. VIZ stores images generated by functions.
The model is setup to run on DLRACD, the cows dataset made from NZ LINZ aerial imagery and MNIST.
This repository primarily uses PyTorch and FrEIA.
You can install the neccesary requirements using the .yml file, e.g. conda env create -f cowflow.yml. FrEIA will need to be installed manually: 'pip install git+https://github.com/VLL-HD/FrEIA.git
'.
This project is licensed under the MIT License.