search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.79k stars 7.96k forks source link

does darknet support multiple types of class？ #2475

Open lqian opened 5 years ago

lqian commented 5 years ago

any possible to detect object with multiple attributes by darknet? for example. detect person's gender, age and clothes color in only one network.

AlexeyAB commented 5 years ago

Yes, Yolo v3 uses multi-label classifications, so you can mark each object with multiple thruth bboxes with different classes, and it will be detected as 1 bbox with multiple classes.

lqian commented 5 years ago

the proposal method neither modify source code nor refine network config file. just make gender label, age label and clothes color label for a person. is it right?

I assume that a ground truth box of a person like this:

0 0.56 0.234 0.15 0.26
10 0.56 0.234 0.15 0.26
20 0.56 0.234 0.15 0.26

AlexeyAB commented 5 years ago

@lqian Yes. Here everything is correct.

Just ensure that you use Yolo v3 cfg file with [yolo] layers instead of Yolo v2 with [region] layers.

lqian commented 5 years ago

does yolov3 support regression age while detecting person?

lqian commented 5 years ago

@AlexeyAB some region always have 0 sample with this method. does it work? i.e. Region 106 Avg IOU: -nan, Class: -nan, Obj: -nan, No Obj: 0.000054, .5R: -nan, .75R: -nan, count: 0 Region 82 Avg IOU: 0.797258, Class: 0.580321, Obj: 0.640600, No Obj: 0.001473, .5R: 1.000000, .75R: 0.750000, count: 8 Region 94 Avg IOU: 0.845069, Class: 0.664617, Obj: 0.723935, No Obj: 0.000782, .5R: 1.000000, .75R: 1.000000, count: 4 Region 82 Avg IOU: 0.853335, Class: 0.618197, Obj: 0.544932, No Obj: 0.001737, .5R: 1.000000, .75R: 1.000000, count: 8 Region 94 Avg IOU: 0.818688, Class: 0.660003, Obj: 0.704443, No Obj: 0.001862, .5R: 1.000000, .75R: 0.909091, count: 22 Region 106 Avg IOU: -nan, Class: -nan, Obj: -nan, No Obj: 0.000018, .5R: -nan, .75R: -nan, count: 0 Region 106 Avg IOU: -nan, Class: -nan, Obj: -nan, No Obj: 0.000004, .5R: -nan, .75R: -nan, count: 0

AlexeyAB commented 5 years ago

@lqian Just there are not objects in your training dataset for these anchors in the last yolo layer.

lqian commented 5 years ago

the model got 93.52% recall after train 23000 iteration customize dataset with a minor modification with yolov3.cfg. thanks a lot!

TinyQi commented 3 years ago

hi, @AlexeyAB .I have some problems that have been bothering me for a long time. Now i know yolov3 can train multi-label classifications. But when I tried to use focal loss to train the model, the results were strange.Every class's prob became to highest level.Another version model without focal loss is ok.Am i do something wrong?

class[1] and class[3] is the right anwser.

Focal Loss Model was weird dets[0] prob list:class[0]:0.990 class[1]:0.991 class[2]:0.988 class[3]:0.992
dets[1] prob list:class[0]:0.997 class[1]:0.998 class[2]:0.996 class[3]:0.998
dets[2] prob list:class[0]:0.991 class[1]:0.992 class[2]:0.988 class[3]:0.993
dets[3] prob list:class[0]:0.997 class[1]:0.998 class[2]:0.995 class[3]:0.998
dets[0] prob list:class[0]:0.997 class[1]:0.998 class[2]:0.997 class[3]:0.998
dets[1] prob list:class[0]:0.997 class[1]:0.998 class[2]:0.996 class[3]:0.998
dets[2] prob list:class[0]:0.996 class[1]:0.996 class[2]:0.995 class[3]:0.997
dets[3] prob list:class[0]:0.997 class[1]:0.998 class[2]:0.996 class[3]:0.998
dets[4] prob list:class[0]:0.998 class[1]:0.998 class[2]:0.996 class[3]:0.998
dets[5] prob list:class[0]:0.997 class[1]:0.997 class[2]:0.995 class[3]:0.998
dets[0] prob list:class[0]:0.826 class[1]:0.827 class[2]:0.826 class[3]:0.827
dets[1] prob list:class[0]:0.995 class[1]:0.995 class[2]:0.994 class[3]:0.996

No Focal Loss Model was right dets[0] prob list:class[0]:0.007 class[1]:0.453 class[2]:0.001 class[3]:0.456
dets[1] prob list:class[0]:0.015 class[1]:0.960 class[2]:0.002 class[3]:0.970
dets[2] prob list:class[0]:0.012 class[1]:0.987 class[2]:0.002 class[3]:0.992
dets[3] prob list:class[0]:0.009 class[1]:0.991 class[2]:0.001 class[3]:0.999
dets[0] prob list:class[0]:0.006 class[1]:0.860 class[2]:0.001 class[3]:0.864
dets[1] prob list:class[0]:0.007 class[1]:0.949 class[2]:0.001 class[3]:0.954
dets[2] prob list:class[0]:0.007 class[1]:0.900 class[2]:0.001 class[3]:0.908
dets[3] prob list:class[0]:0.005 class[1]:0.998 class[2]:0.001 class[3]:0.999
dets[4] prob list:class[0]:0.006 class[1]:0.997 class[2]:0.001 class[3]:0.999
dets[5] prob list:class[0]:0.006 class[1]:0.996 class[2]:0.001 class[3]:0.998
dets[0] prob list:class[0]:0.010 class[1]:0.665 class[2]:0.001 class[3]:0.673