WiderFace accuracy about 62%, not 90%

[borisgribkov]

Hi! Thanks for the provided code first of all!
I guess there is a mistake in "readme.md" file, you mentioned about 90% wider face dataset accuracy, but I tested caffemodel in this repository, model in https://github.com/lippman1125/faceboxes_lqy/tree/ssd/examples/faceboxes/demo and finally trained my own model. Max accuracy reached on WiderFace was about 62% (VOC metric), at the same time it has shown about 90% (VOC) for FDDB with 1024 input.