The project goal was to create a classifier that can distinguish between tree images and non-tree images (like bushes/grass/overlapping trees). A deep neural network was trained over a hand curated dataset and results were fair. In the second round of learning, lessons were employed and a new dataset was curated at even higher quality. The model hyper-parameters were also adjusted and the results were fantastic! 0.98 accuracy!
- Data was dirty making it impossible to run off-the-bat.
- Data collection from multiple sources.
- Investigating why the network doesn’t work, how to test it, how to improve the quality of the data.
Achievements (according to KPIs)
- A classifier model was trained and tested. Accuracy on a balanced binary dataset reached 0.98!
- Visualization and statistics were produced using the model’s output to gain insights on where the model works and where it needs to improve.