cedrickchee / ssd-yolo-retinanet

Multi-class object detection pipeline—Single Shot MultiBox Detector (SSD) + YOLOv3 (real-time) + focal loss (RetinaNet) + Pascal VOC 2007 dataset

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

OHEM and focal loss, which is better?

chengyu0910 opened this issue · comments

Hi, is the focal loss works in SSD? How much improvement is reached in VOC2007 evaluation when SSD is trained with focal loss? I want to know whether the focal loss performs better than OHEM(used in original SSD). Thank you!

For me, focal loss didn't work well in SSD.

I have a question : why exclude the number of negative sample in multi_box_loss? (Line 68 69)