mfwos/Digital-Africa-Plantation-Counting-Challenge

# Introduction The Digital Africa Plantation Counting Challenge is a challenge that was hosted by Zindi (a community of African Data Scientists hosting a wide range of machine learning challenges) and Digital Africa (digital-africa.co). The subject of the challenge was to count the number of palm oil tree in drone images collected from farms in Côte d'Ivoire. This project uses the provided dataset as an example to show how to create a simple baseline model to count instances of a particular object within an image. Further details about the competition can be found here: zindi.africa. This project uses the challenge and its data to first create a simple baseline counting model using a popular pretrained backbone and later try to improve on this baseline by introducing a second model that is trained to boost the initial predictions while using a different resolution scale from the first model. # Preparations of Code and Data As the dataset is property of the competition hosts and publication is not permitted, the project does not contain any actual images nor metadata of the images. The trained models contain a large number of weights and are memory consuming (~50 MB each). Due to this reason the models have not been uploaded to this GitHub page. However, by visiting the competition's homepage and accepting the terms, private access to the dataset can still be gained and the code of this repository can be used to recreate the models and results. To do so, the folder structure of this repository must be copied and the images saved in './data/TreeImages', the (empty) file './data/TreeImages.csv' must be replaced by the real training csv file. The script 'simple_boosting.py' can then be run to create baseline and boosted models for different training and validation splits. To recreate the results presented here all random seeds in 'simple_boosting.py' must stay the same and the script must …