Focal-Loss-implement-on-Tensorflow

This is a multi-label version implementation(unofficial version) of focal loss proposed on Focal Loss for Dense Object Detection by KM He. After implement focal loss formular I have tested on SSD_MobileNet Network on COCO datasets. The following is about the tensorboard results and analysis:

1. The [email protected] of focal loss implementation(num_hard_example=20000, max_neg_per_pos=1000, max_total_detection=20000) are about 3% higher than original model(num_hard_example=2000, max_neg_per_pos=100, max_total_detection=100), about 29% and 26% respectively.

 2. However, why do focal loss can achive so much improvement? Here we show some evaluation images and then we may find the answer. Focal loss do not detect the low percentage object(about 0.5~0.8), this may have a good precision but bad recall. But the core viewpoint of the paper is focus on the hard examples while ignore the easy examples. I don't know the reason. Can anybody tell me the reason???

Conclusion: In my opinion, maybe Focal loss can have a better precision(MAP) but worse recall.
Application: If you just want to have a better MAP, the focal loss is a good choice.