Buildings detection using "detectron2" model developed by META

Roboflow is a end-to-end computer vision tool, and we used it in our work.

Architecture

1. This project aims to detect buildings from the satellite images.

2. The architecture used is called GeneralizedRCNN (Region-Based Convolutional Neural Network).

3. This architecture conists of three blocks

   Feature Pyramid Network (FPN)

   Region Proposal Network (RPN)

   ROI Heads

Detailed explination of each block is present in "https://medium.com/@hirotoschwert/digging-into-detectron-2-47b2e794fabd".

Dataset

The dataset we considered is 'building-detection Image Dataset' from Roboflow.

Code

code is available in the build-seg.ipynb file.

Adjust the file paths before executing

Results

Results are available in the results folder.