search

baomingwang / MTCNN-Tensorflow

mtcnn-tf
MIT License
162 stars 96 forks source link

performance of pretrained model #1

Open RunningLeon opened 6 years ago

RunningLeon commented 6 years ago

Hi, Have you evaluated the pretrained model on any other dataset? Can you post it? Thanks a lot.

baomingwang commented 6 years ago

I have tested it on Wider Face dataset with the official evaluation tools, and the result is about 3~5% lower than the original caffe model. I think the author may have adopted some other training strategy like preparing more negative data, because I noticed that his PNET would generate less false candidates than mine. Maybe you can try something like that.

Good luck!

choice17 commented 6 years ago

@wangbm are you referring to pretrained/*.npy ? or the models at save_model/all_in_one ? or are they actually the same ?