https://github.com/charlie-wartnaby/zindi-arm-unicef-dvc

Zindi Arm UNICEF Disaster Vulnerability Challenge
https://github.com/charlie-wartnaby/zindi-arm-unicef-dvc

Last synced: 7 days ago
JSON representation

Zindi Arm UNICEF Disaster Vulnerability Challenge

• Host: GitHub
• URL: https://github.com/charlie-wartnaby/zindi-arm-unicef-dvc
• Owner: charlie-wartnaby
• License: mit
• Created: 2024-03-28T17:14:58.000Z (8 months ago)
• Default Branch: main
• Last Pushed: 2024-05-19T16:51:35.000Z (6 months ago)
• Last Synced: 2024-08-02T15:30:31.658Z (3 months ago)
• Language: Python
• Size: 288 KB
• Stars: 7
• Watchers: 1
• Forks: 2
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# zindi-arm-unicef-dvc

## Zindi Arm UNICEF Disaster Vulnerability Challenge

This is my entry for this competition:

https://zindi.africa/competitions/arm-unicef-disaster-vulnerability-challenge

This is a fairly straightforward object detection and classification task,

here using YOLOv8. That already does a lot of training data augmentation

for us.

I have explored the augmentation options and some key hyperparameters to try

and get optimal results. Of particular importance are:

* conf: Confidence threshold; trade-off between false negatives and false

        positives.

* iou: Intersection over Union threshold for Non-Maximal Suppression, i.e.

    rejecting lower-confidence detections which overlap a stronger detection

    to some degree. As houses don't overlap (unlike some other probems where

    we may have a dog behind a cat say), this is set fairly low.