Collaborative dataset generation for object detection on satellite imagery

Teacher: Juan B. Pedro — Starlab

Some studies suggest that most data scientists spend only 20% of their time on actual data analysis and 80% of their time finding, cleaning and reorganizing data. This inefficiency can increase even more when working with EO data, since it has a difficult access and it is very expensive to label in terms of time and human expertise. Some datasets exist today for training ML/DL models on EO data, but due to its nature, they are limited to specific tasks on very limited areas.

The POINTOUT project goal is to alleviate this problem by providing easy access to EO data and tools to perform collaborative labeling. Through a web based platform, users can annotate objects directly onto a map to build datasets that can then be downloaded and used to train learning models. We believe that tools like POINTOUT can result in massive speedup on the EO data scientist workflow.

In this session, attendees will be given the task to train an object recognition model from scratch. To that end, they will have to use the POINTOUT platform to download datasets, adding new annotations to existing ones or even creating new datasets with new labels in a collaborative way.

Prerequisites: Google Colaboratory.

#DL  #ML  #AI4EO #EarthObservation #Datascience