Building Footprint Extraction from Remotely Sensed Very High-Resolution Images using Deep Learning
Budget: $250 – $750 USD
Building an automated tool to extract building footprints from remotely sensed images is the main objective of this project. The first step toward this objective will be to build a semantic segmentation model. When it comes to training a deep learning model, there are two distinct approaches:
1. Training a model from scratch
2. Finetuning an existing model using a new dataset
The second approach has become very popular, particularly with datasets with a limited number of label examples. The following approach will be used. There are some research papers and detailed tutorials on finetuning semantic segmentation models. Our basis will be an existing pre-printed model. We will be fine-tuning that pre-trained model using our custom dataset.
The second step of the project will be converting semantic segmentation results into vector maps. Finally, we will be fixing errors in the generated vector maps, and we will further fine-tune these vector maps.
1. Training a model from scratch
2. Finetuning an existing model using a new dataset
The second approach has become very popular, particularly with datasets with a limited number of label examples. The following approach will be used. There are some research papers and detailed tutorials on finetuning semantic segmentation models. Our basis will be an existing pre-printed model. We will be fine-tuning that pre-trained model using our custom dataset.
The second step of the project will be converting semantic segmentation results into vector maps. Finally, we will be fixing errors in the generated vector maps, and we will further fine-tune these vector maps.