Open anguoyang opened 6 years ago
@ sorry for delayed replay, there is no such file, just folder with person.jpg, eg: data/facedb/name1/aligned_face.jpg data/facedb/name2/aligned_face.jpg ..., as for named facedb, cause i have another python project i use sqlite stored my own face data. it has nothing to do with this project.
hi@tpys, thank you for your answer, I have tried your demo, the detection is pretty good, but always display "unknow", anything wrong with me? thank you.
my images in the facedb is the latest photo of myself, and I use my usb camera for testing, I tried to print out the log in your source code, the similarity is always under 0.4, which is lower than the threshold 0.6, I dont know why.
@anguoyang how did you get your latest photot, through 'S' button through when runing this program, if not try this way, it will save the aligned face of yourself through webcam. if you already did this, still get low score, i guess this model doesn't work so well in your environment, I can't help much further.
tpys: Thanks for your contribution. And when i press 'S' button, i only get an empty aligned_face.jpg file. I also met similar problems of very low score, any must i make the alighned face with some special means? Thanks.
Hello every one! I have found the reason why the program on my computer all ways can't recognize the faces. To recognize the faces , the key points are below: 1 Establish subfolders such like Jobs/Messi/Gates in the facedb folder, and place a face photo of Jobs/Messi/Gates in each subfolder.And all the photos must be named with "aligned_face". 2 The photo must be 96(w) x 112(h).96 and 112 is the number I found in the face-everthing code. I select frontal faces of 6 persons and test the program with a threshold of 0.4.It does work! The photos I selected for test are from the sphere face project lfw-112x96 folder.I replaced one of the person's aligned_face with an 114 x 145 picture and the program can't recognize this face.So it seems that size of the aligned_face must be 96 x 112 or keep the width/height ratio be 96/112(I guess ,without any test).All the aligned_face are frontal faces.I guess that faces in profile in the facedb may cause a performance degradation.However, it doesn't matter for the faces in profile in test set, they are all recognizable.I also make two aligned_face of 112x96, both can be recognized.
@macheski Great. Please consider my comment in https://github.com/tpys/face-everthing/issues/6
Hi, guys, thank you so much for your great contribution, could please share the files under the data directory? especially the facedb file, I suppose that it is a sqlite file? thank you.