Use feature pyramid to get image features at different scales. 1) Use top down connection to provide lower layers with high level features. 2) Efficient because all levels use the same backbone and
## 2. Use Focal loss to handle imbalance classes
## 3. Use three separate losses to train the network
