My Team Solution to The Road Segment Identification Hackathon by #ZindiWeekendz
Governments around the world spend resources mapping roads in their countries to make sure that resources, such as emergency services and education, can reach as many people as possible. Many roads are constructed by the government, but many are created by people trying to reach a new location or creating a shortcut. New roads often pop up around development sectors such as mines and farms. In these cases, government officials need to physically confirm the presence of the road, measure and map the road, and ensure that it is usable.
The objective of this competition is to identify whether an image contains a road segment or not. Dry river beds, railway tracks and power lines could look like roads. It is important to classify these as “not roads”.
Our solution was based on a pretrained DenseNet 50 model, and on creating new data using data augmentation.
Final Score(AUC score): 0.9343 (competition winner Score: 0.9699)
Final rank: 38 out of 70