RetinaNet applied to widerface pedestrian detection challenege in ECCV 2018, in PyTorch

This code placed 21 position among 168 teams.

Improvement ideas (todos):

1. Run for more epoch
2. User More data augmentation methods
3. optimize hyper-parameter like learning rate, learning rate decay
4. Use ADAM, SGD with momentum
5. use soft nms
6. use multi scale testing.
7. Optimize FPN feature extractor for small pedestrian object.
8. Use GAN to generate training data for in context of road and pedestrian