Skip to content

Latest commit

 

History

History
21 lines (14 loc) · 1.73 KB

DATASET.md

File metadata and controls

21 lines (14 loc) · 1.73 KB
Raw

Dataset Structure

The project data needs to be structured as specified in this document.

All the data should be placed within the data folder. (i.e. ./data)

Each dataset should be placed in a subfolder under data with a unique name. (e.g. ./data/Sint-Maarten-2017)

Each dataset should have four subfolders within,

  • train containing the images from which the model will learn to identify damage. (typically 80% of all labelled data) (i.e. ./data/Sint-Maarten-2017/train)
  • validation containing the images which will be used to tune the model. (typically 10% of all labelled data) (i.e. ./data/Sint-Maarten-2017/validation)
  • test containing the images which will be used to score the model. (typically 10% of all labelled data) (i.e. ./data/Sint-Maarten-2017/test)
  • inference containing the images for which the model will be used on. (typically all unlabelled data) (i.e. ./data/Sint-Maarten-2017/inference)

Each subfolder is split into two folders and one file,

  • before containing the images of the region before the disaster. (i.e. ./data/Sint-Maarten-2017/test/before)
  • after containing the images of the region after the disaster. (i.e. ./data/Sint-Maarten-2017/test/after)
  • The before and after folders should contain the matching pairs files with the same filename. (e.g. In Sint-Maarten-2017 the building OBJECTID is used as the filename) (i.e. ./data/Sint-Maarten-2017/test/before/10134.png and ./data/Sint-Maarten-2017/test/after/10134.png)
  • Exempted for Inference Set - A text file (labels.txt) with each line containing the filename (before and after images share this name) and the level of damage of the building in this image (i.e. 10134.png 0.6030)