search

Seanlinx / mtcnn

595 stars 264 forks source link

image annotations flip? #31

Open Cv9527 opened 7 years ago

Cv9527 commented 7 years ago

@Seanlinx 152 line in imdb.py: m_bbox[0], m_bbox[2] = -m_bbox[2], -m_bbox[0]. It seems the box is fliped. does the flip is necessary when training model?

Seanlinx commented 7 years ago

@Cv9527 Not necessary, but adding mirrored images will enlarge the training set and improve the performance.

Cv9527 commented 7 years ago

during training Pnet, for the face class task, how to ignore part label -1 when calculate softmaxloss, for regression task, how to ignore negative data?

Seanlinx commented 7 years ago

@Cv9527 The negativemining layer sets the out_grad to "0" where it should be ignored.