A question about the recall of PNet

[LiChenyang-Github]

According to link, the recall should be 97%~99% on FDDB dataset. In my case, I train the MTCNN on wider face and then calculate the recall of PNet on FDDB and the trainset of wider face. On FDDB, the recall (iou thred=0.5) is around 97%. However, on wider face (trainset), the recall is only around 47.8%(iou thred=0.65) or 69.3%(iou thred=0.4). Has anyone calculated the recall of PNet on wider face?